SOV-Transcribe-1
by salesforce-sovereign-intelligence
Salesforce's open-weight speech recognition model: a Whisper large-v3 encoder feeding Gemma 4 26B A4B, Apache 2.0, the most accurate open-weight model on Artificial Analysis
salesforce-sovereign-intelligence/SOV-Transcribe-1Deploy SOV-Transcribe-1
Single-tenantMixpeek has no managed extractor for this model. On a single-tenant deployment you upload the weights and a custom plugin serves them next to the rest of your pipeline.
Overview
SOV-Transcribe-1 is an open-weight speech recognition model from Salesforce, published on Hugging Face on 10 November 2026 under Apache 2.0. It transcribes recordings of up to 30 minutes in English, French, German, Spanish, Italian, Portuguese and Dutch, and takes a plain-language instruction that sets the style (verbatim or cleaned up, number formatting, punctuation) and can list names or terms to listen for. Artificial Analysis ranks it the most accurate open-weight speech-to-text model, at 2.3% AA-WER.
Architecture
A Whisper large-v3 audio encoder, an audio projector that stacks 4 encoder frames into one token, and the Gemma 4 26B A4B mixture-of-experts language model as the decoder: 26.6B parameters in total, about 4.6B active per token, in BF16 with the MoE router in FP32. Audio is 16 kHz mono, padded to 30-second windows of 375 tokens each. It is non-streaming.
Mixpeek SDK Integration
# SOV-Transcribe-1 runs on your side, here behind vLLM with the plugin from the
# model repo (vllm serve ... --max-model-len 32768). Its transcript is upserted into
# a collection whose text vector index makes the recording searchable.
from openai import OpenAI
import requests
asr = OpenAI(base_url="http://127.0.0.1:8000/v1", api_key="EMPTY")
transcript = asr.audio.transcriptions.create(
model="salesforce-sovereign-intelligence/SOV-Transcribe-1",
file=open("q3-earnings-call.wav", "rb"),
prompt="This audio may contain the following words or phrases: Mixpeek, MVS, ARR. Transcribe the audio into text.",
temperature=0,
).text
requests.post(
"https://api.mixpeek.com/v1/namespaces/ns_your_namespace/documents/upsert",
headers={"Authorization": "Bearer API_KEY"},
json={"collection_id": "col_your_collection",
"documents": [{"document_id": "q3-earnings-call",
"vectors": {"text-embedding": your_embedding_of(transcript)},
"payload": {"transcript": transcript, "source": "s3://recordings/q3-earnings-call.wav"}}]},
)Capabilities
- Transcription of recordings up to 30 minutes (tested), one request per file
- Seven languages: English, French, German, Spanish, Italian, Portuguese and Dutch
- Style set by instruction: verbatim with fillers, or cleaned-up written form
- Context biasing with a list of names, terms or a short topic description
- OpenAI-compatible transcription endpoint through vLLM
Use Cases on Mixpeek
Benchmarks
| Dataset | Metric | Score | Source |
|---|---|---|---|
| AA-WER v2 (non-streaming) | Word error rate | 2.3% | Artificial Analysis speech-to-text leaderboard, November 2026: the most accurate open-weight model of 16 evaluated |
| Artificial Analysis overall | Rank, non-streaming | #7 | Model card: salesforce-sovereign-intelligence/SOV-Transcribe-1, as of 12 November 2026 |
| VoiceCodeBench | Entity accuracy and task success | #2 of 23 systems | Model card, as of 12 November 2026 |
Performance
We have not measured it. The speed figures are Salesforce's, from the model card. The card recommends greedy decoding (temperature 0) and notes that in rare cases the model repeats a character or phrase until it reaches the token limit, so check long outputs for loops.
Common Pipeline Companions
Another open-weight transcription model with a mixture-of-experts decoder, to compare on your own audio
Speaker turns, since SOV-Transcribe-1 returns text without speaker labels
Embed the transcripts for retrieval
Frequently Asked Questions
What is SOV-Transcribe-1?
An open-weight speech recognition model from Salesforce, published on Hugging Face on 10 November 2026 under Apache 2.0. It pairs a Whisper large-v3 audio encoder with the Gemma 4 26B A4B language model, transcribes recordings of up to 30 minutes in seven languages, and follows plain-language instructions for transcript style and vocabulary.
How accurate is SOV-Transcribe-1?
Artificial Analysis lists it as the most accurate open-weight speech-to-text model, at 2.3% AA-WER, and the model card puts it at #7 overall on that leaderboard as of 12 November 2026. For comparison, Whisper large-v3 scores 4.07% and ElevenLabs Scribe v3, the most accurate API on our [speech-to-text list](/curated-lists/best-speech-to-text-apis), scores 1.95%.
What languages does SOV-Transcribe-1 support?
English, French, German, Spanish, Italian, Portuguese and Dutch. For anything else, Whisper large-v3 covers 99 languages.
What hardware does SOV-Transcribe-1 need?
It has 26.6B parameters in BF16, so plan on a large GPU. Salesforce's figures are from one RTX PRO 6000 Blackwell with 96 GB, and the card says to set max_num_seqs to 64 with the speculative-decoding draft so vLLM's start-up memory profiling fits in 96 GB. Native vLLM and Transformers support is being upstreamed; until then a plugin in the repo adds it to vLLM 0.32.0.
Does SOV-Transcribe-1 stream?
No. It is a non-streaming model: it transcribes a finished recording. For live captions or voice agents, use a streaming model such as Deepgram Nova-3 or ElevenLabs Scribe v3 Live.
Can I search SOV-Transcribe-1 transcripts with Mixpeek?
Yes. Run the model on your own GPU, embed each transcript, write it to a collection through the documents upsert endpoint with the transcript in the payload, and search it with a retriever. If a long recording comes back with a repeated phrase, see [why speech models make up words during silence](/guides/why-does-whisper-make-up-words-during-silence) for filters that catch loops.
Specification
Build a pipeline with SOV-Transcribe-1
Add this model to a processing pipeline alongside other extractors. Combine with retrieval stages for end-to-end search.
Run it on your own data, free