NEWVectors or files. Pick a path.Start →
    Back to All Lists

    Best Speech-to-Text APIs in 2026

    What is the best speech-to-text API? For recorded audio, ElevenLabs Scribe v3 has the lowest error rate of the major APIs on Artificial Analysis's independent benchmark (1.95%) at $0.20 an hour. For live voice agents and captions, Deepgram streams Nova-3 from $0.0048 a minute. We compare 12 APIs on accuracy, latency, languages and price, checked in November 2026.

    Ethan Steininger
    Published · Updated
    Last tested: July 22, 2026
    12 tools evaluated

    Point Mixpeek at the bucket that holds your audio or video. It writes a transcript in timestamped segments and indexes it, so a search for a line returns the recording and the second it was said. Audio search with transcripts is $0.034 a minute at public rates (November 2026).

    Transcribe your recordings and search them by what was said

    No account needed: Search 69 movie trailers by a line of dialogue

    Key Takeaways

    • ElevenLabs Scribe v3 is our #1 pick for recorded audio. On the Artificial Analysis speech-to-text leaderboard it scores 1.95% word error (AA-WER v2), second of the 65 models it has evaluated and the lowest of any API on this list that it measures. ElevenLabs lists it at $0.20 per hour of audio across 90+ languages.
    • Deepgram is the pick for live audio. Deepgram's pricing page lists Nova-3 streaming from $0.0048 a minute and pre-recorded from $0.0043, and Artificial Analysis measured Nova-3 at 426.8x real time, the second fastest model it tested. Its accuracy trails the leaders: Nova-3 scores 5.18% AA-WER.
    • AssemblyAI and OpenAI sit between those two on accuracy. AssemblyAI's Universal-3.5 Pro scores 2.99% AA-WER at $0.21 an hour for 18 languages, and Universal-2 covers 99 languages at $0.15 an hour (pricing). OpenAI's gpt-transcribe scores 3.31% at $0.0045 a minute, and its whisper-1 and gpt-4o-transcribe endpoints are now marked deprecated on OpenAI's pricing page.
    • To self-host, Whisper large-v3 is MIT-licensed, covers 99 languages and scores 4.07% AA-WER. Put a voice activity filter in front of it: Whisper is known to write phrases nobody said during silence, and its default settings keep a made-up line when the model is confident in it. Why Whisper makes up words during silence shows the fix.
    • Google Cloud Speech-to-Text still covers the most languages, 125+ with dialects. Its V2 API costs $0.016 a minute for the first 500,000 minutes a month (pricing), and Chirp 3 scores 4.32% AA-WER.
    • A transcript is the first step if you want to search what was said. Mixpeek reads your audio and video where they already live, in S3, GCS or any bucket, writes timestamped transcript segments and indexes them, so a search for a line returns the recording and the second it was said. Audio search with transcripts is $0.034 a minute at public rates (pricing).

    Skip the comparison? Mixpeek runs speech-to-text on your own data: extraction, indexing, and search in one platform.

    How We Evaluated

    Evaluated by the Mixpeek engineering team, who build and operate multimodal retrieval infrastructure in production. Last tested July 2026; rankings re-checked when the market shifts, with pricing and claims verified against each vendor's public documentation.

    Accuracy

    30%

    Word error rate across clean speech, noisy environments, accented speakers, and domain-specific terminology.

    Real-Time Performance

    25%

    End-to-end latency for streaming transcription and responsiveness to speech-to-text conversion.

    Language Support

    25%

    Number of supported languages, dialect handling, and accuracy on non-English content.

    Advanced Features

    20%

    Speaker diarization, punctuation, PII redaction, and custom vocabulary support.

    Quick answer

    The short version, before the detail:

    • ElevenLabs Scribebest for teams that want the highest transcription accuracy across many languages with built-in diarizationThe lowest measured error rate of the major APIs (1.95% AA-WER) at the lowest list price for recorded audio ($0.20 an hour), across 90+ languages.
    • Deepgrambest for live transcription and voice agents where latency and per-minute price matter mostAmong the fastest models measured (426.8x real time) with low-latency streaming at low per-minute prices, which makes it the default choice for voice agents and live captions.
    • OpenAI Whisperbest for multilingual transcription, noisy audio, and self-hosted deployments99-language coverage under the MIT license with a large tooling ecosystem, so you can run accurate multilingual transcription on your own infrastructure with no per-minute fee.
    • AssemblyAIbest for developers wanting transcription plus content safety, entity detection, and summarizationThe richest audio intelligence feature set in a single API: PII redaction, content safety, entity detection, sentiment, and LLM-powered summarization without integrating separate services.
    • Google Cloud Speech-to-Textbest for global apps needing 125+ languages, medical transcription, or gcp integrationThe widest language and dialect coverage (125+) combined with specialized domain models (medical, phone, short queries) that no competitor matches.
    • AWS Transcribebest for aws-native teams needing medical transcription or call center analyticsDeepest AWS ecosystem integration (S3 triggers, Lambda, Contact Lens, Comprehend) making it the path of least resistance for AWS-native architectures.
    • Rev AIbest for applications processing conversational speech with accented speakers where accuracy matters more than costModels trained on human-corrected transcripts give Rev AI an edge on conversational speech, accented English, and multi-speaker scenarios that trip up competitors.
    • Speechmaticsbest for global enterprises needing accurate transcription across english accents and european languages with on-prem optionsBest-in-class accuracy on non-US English accents and European languages, with on-prem deployment available to mid-market customers: not gated behind enterprise-only agreements.

    What a speech-to-text API is

    A speech-to-text API takes an audio or video file, or a live audio stream, and returns the spoken words as text, usually with a timestamp for each word or segment. Most also return speaker labels (diarization), punctuation and number formatting, and many sell extras such as PII redaction, keyword boosting or summaries. The APIs work in two modes. Pre-recorded transcription, also called batch or async, takes a finished file and returns the transcript once the whole file is processed. It costs less, and the model sees the full context. Streaming transcription returns partial text while the person is still talking, which live captions and voice agents need, and it costs more: Deepgram charges $0.0048 a minute to stream Nova-3 and $0.0043 for a recorded file. Accuracy is measured as word error rate (WER), the share of words substituted, dropped or inserted compared with a human transcript. Lower is better, and the number depends on the audio. Artificial Analysis's AA-WER v2 averages about 8 hours from three sets (agent conversations, European Parliament speech and earnings calls), chosen for accents, domain vocabulary and hard recording conditions. Test on a sample of your own recordings before you commit, because a model that leads on earnings calls can trail on noisy field audio. Price is quoted per minute or per hour of input audio. Standard models on this list run from $0.15 an hour (AssemblyAI Universal-2) to about $1.44 an hour (AWS Transcribe standard), and diarization, redaction and custom vocabulary are often billed on top.

    Best Speech-to-Text APIs: comparison at a glance

    #ToolBest forPricingKey differentiatorMain limit
    1ElevenLabs ScribeTeams that want the highest transcription accuracy across many languages with built-in diarizationScribe v3 $0.20 per hour of recorded audio; Scribe v3 Live $0.22 per hour (November 2026)The lowest measured error rate of the major APIs (1.95% AA-WER) at the lowest list price for recorded audio ($0.20 an hour), across 90+ languages.No self-hosted option, so audio must go to the ElevenLabs cloud
    2DeepgramLive transcription and voice agents where latency and per-minute price matter mostNova-3 pre-recorded $0.0043/min pay as you go, $0.0036/min on Growth; streaming from $0.0048/min; $200 free credit (November 2026)Among the fastest models measured (426.8x real time) with low-latency streaming at low per-minute prices, which makes it the default choice for voice agents and live captions.5.18% AA-WER on Nova-3, behind ElevenLabs, AssemblyAI and OpenAI on accuracy
    3OpenAI WhisperMultilingual transcription, noisy audio, and self-hosted deploymentsFree self-hosted (MIT). Hosted: gpt-transcribe $0.0045/min, gpt-live-transcribe $0.017/min; whisper-1 and gpt-4o-transcribe $0.006/min, marked deprecated (November 2026)99-language coverage under the MIT license with a large tooling ecosystem, so you can run accurate multilingual transcription on your own infrastructure with no per-minute fee.Self-hosted requires a GPU (large-v3 needs about 10 GB of VRAM)
    4AssemblyAIDevelopers wanting transcription plus content safety, entity detection, and summarizationUniversal-2 $0.15/hr (99 languages); Universal-3.5 Pro $0.21/hr (18 languages); real-time from $0.45/hr; diarization $0.02/hr and other add-ons extra (November 2026)The richest audio intelligence feature set in a single API: PII redaction, content safety, entity detection, sentiment, and LLM-powered summarization without integrating separate services.Advanced features (diarization, PII, sentiment) are billed on top of the base rate
    5Google Cloud Speech-to-TextGlobal apps needing 125+ languages, medical transcription, or GCP integrationV2 standard $0.016/min for the first 500,000 minutes a month, falling to $0.004/min past 2 million; V1 standard $0.024/min; medical $0.078/min (November 2026)The widest language and dialect coverage (125+) combined with specialized domain models (medical, phone, short queries) that no competitor matches.Chirp 3 scores 4.32% AA-WER, behind ElevenLabs, AssemblyAI and OpenAI
    6AWS TranscribeAWS-native teams needing medical transcription or call center analyticsStandard from $0.024/min; Medical $0.075/min; volume discounts availableDeepest AWS ecosystem integration (S3 triggers, Lambda, Contact Lens, Comprehend) making it the path of least resistance for AWS-native architectures.Base accuracy behind Deepgram and Whisper in benchmarks
    7Rev AIApplications processing conversational speech with accented speakers where accuracy matters more than costAsync from $0.02/min; streaming from $0.035/min; volume discounts availableModels trained on human-corrected transcripts give Rev AI an edge on conversational speech, accented English, and multi-speaker scenarios that trip up competitors.English-primary: multilingual support limited to ~15 languages
    8SpeechmaticsGlobal enterprises needing accurate transcription across English accents and European languages with on-prem optionsPay-as-you-go from $0.017/min; on-prem licensing availableBest-in-class accuracy on non-US English accents and European languages, with on-prem deployment available to mid-market customers: not gated behind enterprise-only agreements.Less name recognition in the US market compared to Deepgram or AssemblyAI
    9Azure Speech ServiceMicrosoft-stack enterprises needing speech recognition integrated with Teams, Dynamics, and Azure infrastructureStandard from $0.016/min (batch); real-time from $0.016/min; custom models from $1.40/model/hrThe only speech API with native Microsoft 365 and Teams integration, plus container-based on-prem deployment: ideal for enterprises already invested in the Microsoft ecosystem.Base accuracy behind Deepgram and Whisper on English benchmarks
    10GiggyTeams transcribing large volumes of recorded audio on a tight budget who can wait for async resultsBatch free (queued, up to 25 minutes per recording); Fast $0.04 and Streaming $0.08 per input audio hour on prepaid creditsBenchmark-level accuracy at the lowest published price, with a free Batch tier for recorded audio.Newer API with a short production track record next to Deepgram, Google or AWS
    11GladiaEuropean companies needing GDPR-compliant transcription with production-grade streaming and diarizationFree tier with 10 hrs/month; paid from $0.0061/minGDPR-compliant EU data residency with code-switching detection: the best option for European companies that need multilingual transcription without sending data outside the EU.Newer entrant with less production track record than Deepgram or Google
    12Picovoice LeopardEdge and IoT applications that require offline speech recognition with no cloud dependencyFree for personal use; commercial from $5/device/year or enterprise licensingThe only production-grade speech-to-text engine designed for fully offline, on-device operation: no internet, no cloud, no per-minute costs.Lower accuracy than cloud models (higher WER than Deepgram or Whisper large)

    The Mental Model That Makes Multimodal Search Click

    1

    ElevenLabs Scribe

    Speech-to-text API from ElevenLabs. Scribe v3 scores 1.95% word error on Artificial Analysis's AA-WER v2 benchmark, the lowest of any API on this list that it measures, and ElevenLabs reports 27% fewer errors than Scribe v2. It covers 90+ languages, returns speaker labels and word-level timestamps, and has a Live version for streaming at $0.22 an hour.

    What Sets It Apart

    The lowest measured error rate of the major APIs (1.95% AA-WER) at the lowest list price for recorded audio ($0.20 an hour), across 90+ languages.

    Strengths

    • +1.95% AA-WER on Scribe v3, the lowest of the major APIs Artificial Analysis measures
    • +90+ languages in one model
    • +Speaker diarization and word-level timestamps; ElevenLabs reports the lowest diarization error it tested
    • +$0.20 an hour for recorded audio, below Deepgram's $0.26

    Limitations

    • -No self-hosted option, so audio must go to the ElevenLabs cloud
    • -Newer to speech-to-text than Deepgram or Google, with a shorter production track record
    • -Best known for voice synthesis, so STT-specific enterprise tooling is still maturing

    Real-World Use Cases

    • •High-stakes transcription (legal, medical, media) where accuracy outweighs per-minute cost
    • •Multilingual transcription that needs both strong non-English accuracy and reliable diarization
    • •Call recordings that need speaker labels (agent vs customer) at high accuracy
    • •Subtitling and captioning workflows that depend on precise word-level timestamps

    Choose This When

    When transcription accuracy is the deciding factor, especially for multilingual or multi-speaker audio where you also need reliable diarization.

    Skip This If

    When audio cannot leave your own infrastructure (self-host Whisper instead) or you need a long enterprise compliance track record.

    Integration Example

    from elevenlabs.client import ElevenLabs
    
    client = ElevenLabs(api_key="YOUR_API_KEY")
    with open("interview.mp3", "rb") as audio:
        result = client.speech_to_text.convert(
            model_id="scribe_v2",
            file=audio,
            diarize=True,
            timestamps_granularity="word",
        )
    print(result.text)
    Scribe v3 $0.20 per hour of recorded audio; Scribe v3 Live $0.22 per hour (November 2026)
    Best for: Teams that want the highest transcription accuracy across many languages with built-in diarization
    Visit Website
    2

    Deepgram

    Speech recognition API built for speed. Artificial Analysis measured Nova-3 at 426.8x real time on pre-recorded audio, the second fastest model it tested, and it streams with low latency for live captions and voice agents. Flux STT is Deepgram's streaming model for voice agents. Nova models support 60+ languages. On accuracy Nova-3 trails the leaders at 5.18% AA-WER. Smart formatting adds punctuation, capitalization and numerals at no extra cost.

    What Sets It Apart

    Among the fastest models measured (426.8x real time) with low-latency streaming at low per-minute prices, which makes it the default choice for voice agents and live captions.

    Strengths

    • +Nova-3 pre-recorded at $0.0043 a minute ($0.26 an hour), streaming from $0.0048 a minute
    • +426.8x real time on Artificial Analysis's speed test, the second fastest model measured
    • +Low-latency streaming with Flux STT for voice agents
    • +Smart formatting included; diarization is $0.0020 a minute

    Limitations

    • -5.18% AA-WER on Nova-3, behind ElevenLabs, AssemblyAI and OpenAI on accuracy
    • -Diarization, redaction and keyterm prompting are billed on top of transcription
    • -On-premises deployment requires an enterprise agreement
    • -Custom models and the Growth discount need a committed plan ($4,000 a year in prepaid credits)

    Real-World Use Cases

    • •Live captioning for webinars and virtual events with sub-250ms latency
    • •Transcribing podcast back-catalogs at 100x real-time for full-text search
    • •Real-time voice agent pipelines where low latency is critical for conversational flow
    • •Call center post-call analytics with speaker diarization and topic detection

    Choose This When

    When you need live transcription with low latency, such as voice agents and live captions, or fast batch processing of large archives at a low per-minute price.

    Skip This If

    When the lowest error rate matters most (ElevenLabs Scribe v3 measures 1.95% against Nova-3's 5.18%) or you need on-premises deployment without an enterprise contract.

    Integration Example

    const { createClient } = require("@deepgram/sdk");
    const deepgram = createClient(process.env.DEEPGRAM_API_KEY);
    
    const { result } = await deepgram.listen.prerecorded.transcribeUrl(
      { url: "https://example.com/audio.mp3" },
      { model: "nova-3", smart_format: true, diarize: true }
    );
    console.log(result.results.channels[0].alternatives[0].transcript);
    Nova-3 pre-recorded $0.0043/min pay as you go, $0.0036/min on Growth; streaming from $0.0048/min; $200 free credit (November 2026)
    Best for: Live transcription and voice agents where latency and per-minute price matter most
    Visit Website
    3

    OpenAI Whisper

    Open-source speech recognition model from OpenAI. Whisper large-v3 was trained on 1 million hours of weakly labeled audio and 4 million hours of pseudo-labeled audio, covers 99 languages and scores 4.07% on Artificial Analysis's AA-WER v2. It is MIT-licensed, so you can run it on your own GPUs. For hosted use, OpenAI's pricing page now marks whisper-1 and gpt-4o-transcribe as deprecated and lists gpt-transcribe ($0.0045 a minute, 3.31% AA-WER) and gpt-live-transcribe ($0.017 a minute, streaming).

    What Sets It Apart

    99-language coverage under the MIT license with a large tooling ecosystem, so you can run accurate multilingual transcription on your own infrastructure with no per-minute fee.

    Strengths

    • +99 languages from one MIT-licensed model
    • +Free to self-host, so audio never leaves your infrastructure
    • +4.07% AA-WER on large-v3, ahead of Deepgram Nova-3 and Google Chirp 3
    • +Large ecosystem: faster-whisper, whisper.cpp, WhisperX for diarization

    Limitations

    • -Self-hosted requires a GPU (large-v3 needs about 10 GB of VRAM)
    • -Can write phrases nobody said during long silences, so run it behind a voice activity filter
    • -No built-in speaker diarization (needs WhisperX or pyannote)
    • -Hosted whisper-1 is marked deprecated; OpenAI's current hosted models are gpt-transcribe and gpt-live-transcribe

    Real-World Use Cases

    • •Transcribing multilingual customer support calls across 99+ languages
    • •Processing field recordings with heavy background noise (construction, outdoor events)
    • •Building a self-hosted transcription pipeline to avoid sending audio to third-party APIs
    • •Academic research requiring reproducible, free-to-use speech recognition

    Choose This When

    When you need strong multilingual support, want to self-host for privacy or cost reasons, or are processing noisy audio where Whisper's robustness shines.

    Skip This If

    When you need real-time streaming transcription or built-in speaker diarization without stitching together additional libraries.

    Integration Example

    import whisper
    
    model = whisper.load_model("large-v3")
    result = model.transcribe(
        "interview.mp3",
        language=None,          # auto-detect
        word_timestamps=True
    )
    for segment in result["segments"]:
        print(f"[{segment['start']:.1f}s] {segment['text']}")
    Free self-hosted (MIT). Hosted: gpt-transcribe $0.0045/min, gpt-live-transcribe $0.017/min; whisper-1 and gpt-4o-transcribe $0.006/min, marked deprecated (November 2026)
    Best for: Multilingual transcription, noisy audio, and self-hosted deployments
    Visit Website
    4

    AssemblyAI

    Speech-to-text platform that goes beyond transcription into audio intelligence, offering speaker diarization, PII redaction, content safety, entity recognition, sentiment analysis, and auto-summarization in one API. Universal-3.5 Pro, its most accurate async model, scores 2.99% AA-WER across 18 languages at $0.21 an hour, and Universal-2 covers 99 languages at $0.15 an hour.

    What Sets It Apart

    The richest audio intelligence feature set in a single API: PII redaction, content safety, entity detection, sentiment, and LLM-powered summarization without integrating separate services.

    Strengths

    • +Audio intelligence suite: PII redaction, safety, entities, sentiment, summaries
    • +Excellent developer experience with strong docs and SDKs in the category
    • +Universal-3.5 Pro at 2.99% AA-WER; Universal-2 covers 99 languages at $0.15 an hour
    • +LeMUR integration for LLM-powered audio Q&A and summarization

    Limitations

    • -Advanced features (diarization, PII, sentiment) are billed on top of the base rate
    • -Cloud-only with no self-hosted deployment option
    • -Universal-3.5 Pro, the most accurate tier, covers 18 languages
    • -Some features (safety, PII) add latency to the processing pipeline

    Real-World Use Cases

    • •Healthcare call transcription with automatic PII redaction for HIPAA compliance
    • •Podcast production workflows that auto-generate show notes, chapters, and summaries
    • •Content moderation for user-generated audio on social platforms
    • •Sales call analysis with entity extraction, sentiment tracking, and action item detection

    Choose This When

    When you need more than raw transcription, such as content moderation, entity extraction, summarization, or PII handling, and want it all from one vendor with strong docs.

    Skip This If

    When you need the lowest error rate (ElevenLabs Scribe v3 measures 1.95% against Universal-3.5 Pro's 2.99%) or fully on-prem deployment, which AssemblyAI does not offer.

    Integration Example

    import assemblyai as aai
    
    aai.settings.api_key = "YOUR_API_KEY"
    transcriber = aai.Transcriber()
    
    config = aai.TranscriptionConfig(
        speaker_labels=True,
        auto_highlights=True,
        entity_detection=True,
        sentiment_analysis=True
    )
    transcript = transcriber.transcribe("meeting.mp3", config=config)
    for utterance in transcript.utterances:
        print(f"Speaker {utterance.speaker}: {utterance.text}")
    Universal-2 $0.15/hr (99 languages); Universal-3.5 Pro $0.21/hr (18 languages); real-time from $0.45/hr; diarization $0.02/hr and other add-ons extra (November 2026)
    Best for: Developers wanting transcription plus content safety, entity detection, and summarization
    Visit Website
    5

    Google Cloud Speech-to-Text

    Google's speech recognition API with the widest language coverage at 125+ languages and dialects. Offers specialized models for medical dictation, phone calls, and short queries. V2 API with Chirp model available on-prem via Google Distributed Cloud.

    What Sets It Apart

    The widest language and dialect coverage (125+) combined with specialized domain models (medical, phone, short queries) that no competitor matches.

    Strengths

    • +125+ languages with dialect-level support: widest coverage available
    • +Specialized models: Medical Conversations, Phone Call, Short Queries
    • +Chirp model available on-device and on-prem
    • +Multi-channel recognition for call center stereo audio

    Limitations

    • -Chirp 3 scores 4.32% AA-WER, behind ElevenLabs, AssemblyAI and OpenAI
    • -Complex pricing across 3+ model tiers (standard, enhanced, chirp)
    • -GCP lock-in for best integration and lowest latency
    • -Speaker diarization less accurate than AssemblyAI for multi-speaker

    Real-World Use Cases

    • •Global SaaS products serving users in 100+ countries with dialect-specific recognition
    • •Medical dictation systems requiring HIPAA-compliant specialized models
    • •IVR and voice assistant systems using the Short Queries model for command recognition
    • •Multi-channel call center recordings with per-channel speaker separation

    Choose This When

    When your application serves a global user base with rare languages, needs medical-grade transcription, or is already deeply integrated with GCP.

    Skip This If

    When accuracy matters most (ElevenLabs, AssemblyAI and OpenAI measure ahead of Chirp 3) or when you want simple, flat-rate pricing.

    Integration Example

    from google.cloud import speech_v2
    
    client = speech_v2.SpeechClient()
    config = speech_v2.RecognitionConfig(
        auto_decoding_config=speech_v2.AutoDetectDecodingConfig(),
        language_codes=["en-US"],
        model="chirp_2",
        features=speech_v2.RecognitionFeatures(
            enable_automatic_punctuation=True
        ),
    )
    with open("audio.wav", "rb") as f:
        response = client.recognize(
            config=config, content=f.read(),
            recognizer="projects/my-project/locations/global/recognizers/_"
        )
    print(response.results[0].alternatives[0].transcript)
    V2 standard $0.016/min for the first 500,000 minutes a month, falling to $0.004/min past 2 million; V1 standard $0.024/min; medical $0.078/min (November 2026)
    Best for: Global apps needing 125+ languages, medical transcription, or GCP integration
    Visit Website
    6

    AWS Transcribe

    Amazon's speech-to-text service supporting 100+ languages with automatic language identification. Includes Transcribe Medical for HIPAA-eligible clinical dictation and Contact Lens integration for call center analytics with sentiment and issue detection.

    What Sets It Apart

    Deepest AWS ecosystem integration (S3 triggers, Lambda, Contact Lens, Comprehend) making it the path of least resistance for AWS-native architectures.

    Strengths

    • +100+ languages with automatic language identification
    • +Transcribe Medical for HIPAA-eligible clinical dictation
    • +Contact Lens integration for call center analytics
    • +Custom vocabulary and custom language models for domain terms

    Limitations

    • -Base accuracy behind Deepgram and Whisper in benchmarks
    • -Per-second pricing expensive for long audio files
    • -Deep AWS dependency: hard to migrate away
    • -Custom language model training requires substantial data

    Real-World Use Cases

    • •S3-triggered transcription pipelines for media assets uploaded to AWS
    • •HIPAA-compliant clinical note dictation in healthcare EHR integrations
    • •Amazon Connect call center analytics with real-time sentiment and issue detection
    • •Subtitle generation for video-on-demand platforms hosted on AWS infrastructure

    Choose This When

    When your infrastructure is on AWS and you want seamless integration with S3, Lambda, and Contact Lens without managing cross-cloud networking.

    Skip This If

    When raw transcription accuracy is paramount (Deepgram and Whisper outperform) or when you need a cloud-agnostic solution.

    Integration Example

    import boto3
    
    transcribe = boto3.client("transcribe")
    transcribe.start_transcription_job(
        TranscriptionJobName="my-job",
        Media={"MediaFileUri": "s3://my-bucket/audio.mp3"},
        MediaFormat="mp3",
        LanguageCode="en-US",
        Settings={
            "ShowSpeakerLabels": True,
            "MaxSpeakerLabels": 5,
            "ShowAlternatives": True,
            "MaxAlternatives": 3,
        },
        OutputBucketName="my-output-bucket",
    )
    Standard from $0.024/min; Medical $0.075/min; volume discounts available
    Best for: AWS-native teams needing medical transcription or call center analytics
    Visit Website
    7

    Rev AI

    Speech-to-text API from Rev, the human transcription company that used millions of hours of human-corrected transcripts to train its ASR models. Offers both streaming and async transcription with strong accuracy on conversational speech, accented English, and multi-speaker audio.

    What Sets It Apart

    Models trained on human-corrected transcripts give Rev AI an edge on conversational speech, accented English, and multi-speaker scenarios that trip up competitors.

    Strengths

    • +Trained on millions of hours of human-corrected transcripts for high accuracy
    • +Strong performance on accented English and conversational speech
    • +Both streaming and async endpoints with simple REST API
    • +Custom vocabulary support for domain-specific terminology

    Limitations

    • -English-primary: multilingual support limited to ~15 languages
    • -Higher latency on streaming compared to Deepgram
    • -No built-in audio intelligence features (PII, sentiment)
    • -Smaller ecosystem and community compared to Whisper or Deepgram

    Real-World Use Cases

    • •Legal deposition transcription where accuracy on conversational speech is critical
    • •Transcribing interviews and focus groups with multiple accented speakers
    • •Media production workflows requiring broadcast-quality transcripts
    • •Accessibility compliance for video content with diverse speaker profiles

    Choose This When

    When you are transcribing conversational content (interviews, depositions, meetings) where speaker diversity and accent handling matter more than per-minute cost.

    Skip This If

    When you need multilingual support beyond English or want built-in audio intelligence features like PII redaction.

    Integration Example

    from rev_ai import apiclient
    
    client = apiclient.RevAiAPIClient("YOUR_ACCESS_TOKEN")
    job = client.submit_job_url(
        media_url="https://example.com/interview.mp3",
        metadata="legal-deposition-2026",
        skip_diarization=False,
        custom_vocabularies=[{
            "phrases": ["habeas corpus", "amicus curiae"]
        }]
    )
    # Poll for completion
    transcript = client.get_transcript_text(job.id)
    print(transcript)
    Async from $0.02/min; streaming from $0.035/min; volume discounts available
    Best for: Applications processing conversational speech with accented speakers where accuracy matters more than cost
    Visit Website
    8

    Speechmatics

    UK-based speech recognition provider with strong global language support (50+ languages) and an emphasis on accuracy across accents and dialects. Their Ursa model delivers competitive WER while offering on-prem deployment for data-sensitive industries.

    What Sets It Apart

    Best-in-class accuracy on non-US English accents and European languages, with on-prem deployment available to mid-market customers: not gated behind enterprise-only agreements.

    Strengths

    • +50+ languages with strong accent and dialect handling
    • +On-premises deployment available without enterprise-only gates
    • +Competitive accuracy on non-US English accents (UK, Australian, Indian)
    • +Real-time and batch processing with consistent API

    Limitations

    • -Less name recognition in the US market compared to Deepgram or AssemblyAI
    • -Smaller ecosystem and fewer community integrations
    • -Higher pricing than Deepgram for comparable English accuracy
    • -Limited audio intelligence features beyond transcription

    Real-World Use Cases

    • •UK government and public sector transcription with on-premises data residency requirements
    • •European multilingual customer support with strong accent handling across EU languages
    • •Broadcasting and media companies needing accurate subtitles for diverse English accents
    • •Financial services compliance recording where data cannot leave the corporate network

    Choose This When

    When your users speak with diverse accents (UK, Indian, Australian English) or European languages, and you may need on-prem deployment for data sovereignty.

    Skip This If

    When you need the lowest error rate (ElevenLabs Scribe v3 and AssemblyAI Universal-3.5 Pro measure lower) or extensive audio intelligence features (AssemblyAI wins).

    Integration Example

    import speechmatics
    
    sm_client = speechmatics.client.WebsocketClient(
        speechmatics.models.ConnectionSettings(
            url="wss://eu2.rt.speechmatics.com/v2",
            auth_token="YOUR_API_KEY",
        )
    )
    conf = speechmatics.models.TranscriptionConfig(
        language="en",
        enable_partials=True,
        operating_point="enhanced",
    )
    sm_client.add_event_handler(
        event_name=speechmatics.models.ServerMessageType.AddTranscript,
        event_handler=lambda msg: print(msg["metadata"]["transcript"]),
    )
    await sm_client.run(audio_stream, conf)
    Pay-as-you-go from $0.017/min; on-prem licensing available
    Best for: Global enterprises needing accurate transcription across English accents and European languages with on-prem options
    Visit Website
    9

    Azure Speech Service

    Microsoft's speech recognition API within Azure Cognitive Services, supporting 100+ languages with custom speech models, real-time and batch transcription, and tight integration with Microsoft 365 and Teams. Offers on-prem deployment via Azure containers.

    What Sets It Apart

    The only speech API with native Microsoft 365 and Teams integration, plus container-based on-prem deployment: ideal for enterprises already invested in the Microsoft ecosystem.

    Strengths

    • +100+ languages with custom speech model training
    • +On-prem deployment via Docker containers
    • +Deep integration with Microsoft 365, Teams, and Dynamics
    • +Pronunciation assessment and custom voice features

    Limitations

    • -Base accuracy behind Deepgram and Whisper on English benchmarks
    • -Complex pricing tiers (standard, custom, real-time, batch)
    • -Azure ecosystem lock-in for optimal performance
    • -Custom model training UI less intuitive than competitors

    Real-World Use Cases

    • •Transcribing Microsoft Teams meetings with speaker identification for enterprise compliance
    • •Building voice-enabled Dynamics 365 workflows for CRM and customer service
    • •On-premises speech processing in air-gapped environments using Azure containers
    • •Custom speech models for industry-specific vocabulary in manufacturing or finance

    Choose This When

    When your organization runs on Microsoft infrastructure and needs speech recognition that plugs directly into Teams, Dynamics, and Azure services.

    Skip This If

    When you need best-in-class accuracy on English (Deepgram leads) or want a vendor-neutral solution not tied to a cloud provider.

    Integration Example

    import azure.cognitiveservices.speech as speechsdk
    
    speech_config = speechsdk.SpeechConfig(
        subscription="YOUR_KEY",
        region="eastus"
    )
    speech_config.speech_recognition_language = "en-US"
    
    audio_config = speechsdk.audio.AudioConfig(filename="meeting.wav")
    recognizer = speechsdk.SpeechRecognizer(
        speech_config=speech_config,
        audio_config=audio_config
    )
    result = recognizer.recognize_once()
    print(result.text)
    Standard from $0.016/min (batch); real-time from $0.016/min; custom models from $1.40/model/hr
    Best for: Microsoft-stack enterprises needing speech recognition integrated with Teams, Dynamics, and Azure infrastructure
    Visit Website
    10

    Giggy

    Low-cost transcription API with a free queued Batch mode and paid Fast and Streaming modes, covering 25+ languages with automatic language detection. In Verli's public benchmark (July 2026, 21 engines, code on GitHub), Giggy's Fast mode placed third on clean English speech at 1.67% WER and tied for second on hard earnings-call audio at 10.14%, at $0.04 per audio hour.

    What Sets It Apart

    Benchmark-level accuracy at the lowest published price, with a free Batch tier for recorded audio.

    Strengths

    • +Accuracy near the top of an independent benchmark: 1.67% WER on LibriSpeech test-clean and 10.14% on Earnings-22 in Verli's July 2026 run
    • +Among the cheapest published rates: Batch is free, Fast is $0.04 per input audio hour, Streaming $0.08
    • +Fast and uploaded-recording Streaming accept recordings up to 12 hours
    • +Automatic language detection, or set the language explicitly per request

    Limitations

    • -Newer API with a short production track record next to Deepgram, Google or AWS
    • -25+ languages, fewer than Google, AWS or Azure
    • -Async only for recordings: you submit a job and poll for the result, which Verli measured at 4.5 seconds per clip
    • -Free Batch takes recordings of 0.5 seconds to 25 minutes and sits in a queue

    Real-World Use Cases

    • •Podcast or lecture archive transcribed in bulk on the free Batch tier
    • •Research team transcribing interviews in several languages with automatic detection
    • •Earnings-call or meeting recordings where accuracy on hard audio matters more than latency

    Choose This When

    When cost per audio hour drives the decision, your audio is recorded rather than live, and you can poll for results.

    Skip This If

    When you need the widest language coverage, synchronous low-latency responses, or a vendor with years of enterprise deployments.

    Integration Example

    curl 'https://api.giggy.ai/v1/audio/transcriptions?mode=batch' \
      -H "xi-api-key: $GIGGY_API_KEY" \
      -H "Content-Type: audio/wav" \
      --data-binary @recording.wav
    # Returns HTTP 202 with a generation_uuid.
    # Poll GET https://api.giggy.ai/v1/generations/{generation_uuid} until it completes.
    Batch free (queued, up to 25 minutes per recording); Fast $0.04 and Streaming $0.08 per input audio hour on prepaid credits
    Best for: Teams transcribing large volumes of recorded audio on a tight budget who can wait for async results
    Visit Website
    11

    Gladia

    European speech-to-text API that wraps Whisper and proprietary models with production-grade features. Offers real-time streaming, speaker diarization, and audio intelligence in a single API, with GDPR-compliant EU data residency.

    What Sets It Apart

    GDPR-compliant EU data residency with code-switching detection: the best option for European companies that need multilingual transcription without sending data outside the EU.

    Strengths

    • +GDPR-compliant with EU data residency options
    • +Combines Whisper accuracy with production features (streaming, diarization)
    • +Code-switching detection for multilingual conversations
    • +Simple per-minute pricing with no feature add-on costs

    Limitations

    • -Newer entrant with less production track record than Deepgram or Google
    • -Accuracy depends on underlying Whisper models: not custom-trained
    • -Smaller language support than Google or AWS
    • -Limited enterprise features (SSO, audit logs) compared to incumbents

    Real-World Use Cases

    • •EU-based SaaS products requiring GDPR-compliant speech processing with data residency
    • •Multilingual meetings where speakers switch between languages mid-sentence
    • •Startups wanting Whisper-level accuracy with managed streaming infrastructure
    • •European media companies needing compliant transcription for broadcast content

    Choose This When

    When GDPR compliance and EU data residency are requirements, or when your audio contains multilingual conversations with code-switching.

    Skip This If

    When you need a battle-tested enterprise platform with extensive compliance certifications beyond GDPR.

    Integration Example

    const response = await fetch("https://api.gladia.io/v2/transcription", {
      method: "POST",
      headers: {
        "x-gladia-key": process.env.GLADIA_API_KEY,
        "Content-Type": "application/json",
      },
      body: JSON.stringify({
        audio_url: "https://example.com/audio.mp3",
        diarization: true,
        language_behaviour: "automatic multiple languages",
      }),
    });
    const { id } = await response.json();
    // Poll for result at /v2/transcription/{id}
    Free tier with 10 hrs/month; paid from $0.0061/min
    Best for: European companies needing GDPR-compliant transcription with production-grade streaming and diarization
    Visit Website
    12

    Picovoice Leopard

    On-device speech-to-text engine that runs entirely locally without any cloud dependency. Optimized for edge deployment on mobile, embedded, and IoT devices with small model sizes and low resource requirements.

    What Sets It Apart

    The only production-grade speech-to-text engine designed for fully offline, on-device operation: no internet, no cloud, no per-minute costs.

    Strengths

    • +Fully on-device: no internet connection required
    • +Low resource footprint suitable for mobile and embedded devices
    • +No per-minute API costs: one-time or annual licensing
    • +Complete data privacy: audio never leaves the device

    Limitations

    • -Lower accuracy than cloud models (higher WER than Deepgram or Whisper large)
    • -Limited language support compared to cloud alternatives
    • -No speaker diarization or advanced audio intelligence
    • -Smaller model means less robustness on noisy audio

    Real-World Use Cases

    • •Offline voice control for IoT devices in warehouses or field environments without internet
    • •Mobile apps that transcribe voice notes without sending audio to the cloud
    • •Automotive in-car voice systems requiring real-time recognition without cellular connectivity
    • •Healthcare point-of-care devices where patient audio cannot leave the device for privacy reasons

    Choose This When

    When your deployment environment has no reliable internet connectivity, when audio data cannot leave the device for privacy reasons, or when you want to eliminate per-minute API costs.

    Skip This If

    When you need cloud-grade accuracy, broad language support, or advanced features like diarization and audio intelligence.

    Integration Example

    import pvleopard
    
    leopard = pvleopard.create(
        access_key="YOUR_ACCESS_KEY",
        model_path="leopard_params.pv"
    )
    transcript, words = leopard.process_file("recording.wav")
    print(transcript)
    for word in words:
        print(f"{word.word} [{word.start_sec:.1f}s - {word.end_sec:.1f}s] "
              f"confidence: {word.confidence:.2f}")
    leopard.delete()
    Free for personal use; commercial from $5/device/year or enterprise licensing
    Best for: Edge and IoT applications that require offline speech recognition with no cloud dependency
    Visit Website

    Which one fits your use case?

    Find your situation in the first column. The last column says when the pick is the wrong call.

    Choose it whenPickLook elsewhere when
    Transcription accuracy is the deciding factor, especially for multilingual or multi-speaker audio where you also need reliable diarization.ElevenLabs ScribeAudio cannot leave your own infrastructure (self-host Whisper instead) or you need a long enterprise compliance track record.
    You need live transcription with low latency, such as voice agents and live captions, or fast batch processing of large archives at a low per-minute price.DeepgramThe lowest error rate matters most (ElevenLabs Scribe v3 measures 1.95% against Nova-3's 5.18%) or you need on-premises deployment without an enterprise contract.
    You need strong multilingual support, want to self-host for privacy or cost reasons, or are processing noisy audio where Whisper's robustness shines.OpenAI WhisperYou need real-time streaming transcription or built-in speaker diarization without stitching together additional libraries.
    You need more than raw transcription, such as content moderation, entity extraction, summarization, or PII handling, and want it all from one vendor with strong docs.AssemblyAIYou need the lowest error rate (ElevenLabs Scribe v3 measures 1.95% against Universal-3.5 Pro's 2.99%) or fully on-prem deployment, which AssemblyAI does not offer.
    Your application serves a global user base with rare languages, needs medical-grade transcription, or is already deeply integrated with GCP.Google Cloud Speech-to-TextAccuracy matters most (ElevenLabs, AssemblyAI and OpenAI measure ahead of Chirp 3) or when you want simple, flat-rate pricing.
    Your infrastructure is on AWS and you want seamless integration with S3, Lambda, and Contact Lens without managing cross-cloud networking.AWS TranscribeRaw transcription accuracy is paramount (Deepgram and Whisper outperform) or when you need a cloud-agnostic solution.
    You are transcribing conversational content (interviews, depositions, meetings) where speaker diversity and accent handling matter more than per-minute cost.Rev AIYou need multilingual support beyond English or want built-in audio intelligence features like PII redaction.
    Your users speak with diverse accents (UK, Indian, Australian English) or European languages, and you may need on-prem deployment for data sovereignty.SpeechmaticsYou need the lowest error rate (ElevenLabs Scribe v3 and AssemblyAI Universal-3.5 Pro measure lower) or extensive audio intelligence features (AssemblyAI wins).
    Managed Mixpeek

    Put speech-to-text to work

    Connect a bucket and Mixpeek runs the whole speech-to-text pipeline for you: extraction, indexing, and search over your own objects. No models to wire up, nothing to host.

    Start with Managed
    MVS · bring your own

    Already have vectors?

    Keep your embeddings on your own cloud and run dense, sparse, and BM25 search directly on object storage. From $25/mo.

    Start with MVS

    Frequently Asked Questions

    What are the best speech-to-text APIs?

    For recorded audio, ElevenLabs Scribe v3 ($0.20 an hour, 1.95% word error on Artificial Analysis's AA-WER v2) and AssemblyAI Universal-3.5 Pro ($0.21 an hour, 2.99%). For live audio and voice agents, Deepgram Nova-3 and Flux, streaming from $0.0048 a minute. For a hosted OpenAI model, gpt-transcribe ($0.0045 a minute, 3.31%). To self-host, Whisper large-v3 (MIT, 99 languages, 4.07%), or for its seven languages, SOV-Transcribe-1 from Salesforce (Apache 2.0, 2.3%, the most accurate open-weight model Artificial Analysis lists). For the most languages, Google Cloud Speech-to-Text (125+). Prices are list prices checked in November 2026.

    Which speech-to-text API is best for real-time voice agents?

    Pick a streaming model with low latency and turn handling. Deepgram streams Nova-3 from $0.0048 a minute and offers Flux STT for voice agents; AssemblyAI's Universal-3.6 Pro Realtime is $0.45 an hour across 32 languages; ElevenLabs Scribe v3 Live is $0.22 an hour; OpenAI's gpt-live-transcribe is $0.017 a minute. Measure latency from your own region with your own audio, since the round trip to the vendor is part of what the caller hears.

    What is the cheapest speech-to-text API?

    Among the major APIs at November 2026 list prices, AssemblyAI Universal-2 is lowest at $0.15 an hour, then ElevenLabs Scribe v3 at $0.20 and Deepgram Nova-3 pre-recorded at $0.26. Artificial Analysis lists cheaper niche providers, down to $0.37 per 1,000 minutes, and self-hosted Whisper costs only GPU time. Check whether diarization is included: Deepgram bills it at $0.0020 a minute and AssemblyAI at $0.02 an hour.

    Why does my transcript contain words nobody said?

    Whisper-family models can write fluent text over silence, music or noise, such as "Thank you for watching". A study of Whisper found that roughly 1% of transcriptions contained entire hallucinated phrases, and that they occurred more often for speakers with longer pauses. The fix is to cut non-speech out before decoding with a voice activity filter and to drop segments the model flags as probable silence. Why Whisper makes up words during silence has the settings and a runnable filter.

    How is a full audio search pipeline different from a speech-to-text API?

    A speech-to-text API returns a transcript. Making audio actually *searchable* takes three more stages these APIs leave to you: **speaker diarization** (who spoke each segment), **forced alignment** (word-level timestamps so a match jumps to the exact moment), and **embedding plus indexing** so an agent can retrieve by meaning, speaker, and time. Mixpeek runs that whole chain over your storage: transcription with models like Qwen3-ASR, speaker turns with pyannote 3.1, and word timing with Qwen3-ForcedAligner, so a query like 'when did the CEO mention pricing' returns the exact clip, the speaker, and the timestamp. The underlying concepts are in the speaker diarization and forced alignment guides.

    What is the best speech-to-text API for accuracy?

    On Artificial Analysis's AA-WER v2 benchmark (November 2026), ElevenLabs Scribe v3 has the lowest error rate of the major APIs at 1.95%, followed by AssemblyAI Universal-3.5 Pro (2.99%), OpenAI gpt-transcribe (3.31%), self-hosted Whisper large-v3 (4.07%), Google Chirp 3 (4.32%) and Deepgram Nova-3 (5.18%). The benchmark mixes agent conversations, European Parliament speech and earnings calls, so test a sample of your own audio before you choose.

    How much does speech-to-text cost per hour of audio?

    At November 2026 list prices: AssemblyAI Universal-2 is $0.15 an hour, ElevenLabs Scribe v3 $0.20, AssemblyAI Universal-3.5 Pro $0.21, Deepgram Nova-3 $0.26 for recorded audio, and OpenAI gpt-transcribe $0.27. Google's V2 API is $0.96 an hour for the first 500,000 minutes a month and AWS Transcribe standard is $1.44. Streaming costs more than recorded audio at most vendors, self-hosted Whisper costs only your GPU time, and diarization and PII redaction are often billed on top.

    Can speech-to-text APIs handle noisy audio?

    Modern APIs handle moderate background noise well, but accuracy degrades significantly in very noisy environments. Whisper is particularly robust to noise. For best results, preprocess audio with noise reduction and choose models trained on noisy data. Expect 5-15% higher word error rates in noisy conditions versus clean speech.

    Do I need a separate platform if I want to search across my transcripts?

    A speech-to-text API gives you the transcript. If your goal is searching across many audio or video files (find the moment someone says X, or retrieve clips that are semantically about a topic), you still have to chunk, embed, index, and rank that text yourself. Multimodal platforms like Mixpeek run transcription as one stage of an indexing and retrieval pipeline over raw objects, so the search layer is handled rather than bolted on. If you only need transcripts for display or one-off processing, a standalone STT API is simpler and cheaper.

    What are the best alternatives to Google Cloud Speech-to-Text?

    It depends on why you are leaving. For accuracy, ElevenLabs Scribe v3 (1.95% AA-WER), AssemblyAI Universal-3.5 Pro (2.99%) and OpenAI gpt-transcribe (3.31%) all measure ahead of Google's Chirp 3 (4.32%) on Artificial Analysis's benchmark. For price, AssemblyAI Universal-2 is $0.15 an hour against Google V2's $0.96. To self-host, OpenAI Whisper large-v3 is MIT-licensed and covers 99 languages. Google still has the widest coverage, 125+ languages and dialects, so a long-tail language is the main reason to stay. To search what was said across a whole library, see audio search with Mixpeek.

    What are the best alternatives to Azure Speech to Text?

    Teams usually switch for English accuracy or simpler pricing. ElevenLabs Scribe v3 (1.95% AA-WER, 90+ languages) and AssemblyAI Universal-3.5 Pro (2.99%) have the lowest error rates of the major APIs on Artificial Analysis's benchmark, and Deepgram Nova-3 is the low-latency choice for live audio. If you need on-premises deployment, which Azure offers through containers, Speechmatics also runs on-prem and Whisper can be self-hosted under the MIT license. For EU data residency, Gladia offers GDPR-compliant hosting in the EU. Azure's integration with Microsoft 365 and Teams is the main reason to keep it.

    See how Mixpeek handles this

    Purpose-built for speech-to-text apis, not bolted on.

    Audio Search

    Mixpeek's dedicated page for this capability: architecture, benchmarks, and how it works.

    Explore Audio Search

    Talk to a Mixpeek engineer: free

    30 minutes. Bring your use case and we'll tell you exactly what would work and what wouldn't.

    Schedule a Free Call

    Explore Other Curated Lists

    content processing

    Best AI Content Moderation Tools

    We evaluated content moderation platforms across image, video, text, and audio moderation. This guide covers accuracy, latency, customization, and compliance features for trust and safety teams.

    9 tools rankedView List
    content processing

    Best Document AI Platforms

    A hands-on evaluation of platforms for intelligent document processing, including OCR, layout analysis, table extraction, and document search. Tested on invoices, contracts, and technical manuals.

    10 tools rankedView List
    content processing

    Best Audio Processing & Search Tools

    An evaluation of platforms for audio transcription, analysis, and search. We tested on podcasts, call recordings, music, and environmental audio across multiple languages.

    9 tools rankedView List