Served on MixpeekTranscriptionRecommendedVerified 2026-11-05
Whisper large-v3-turbo
openai/whisper-large-v3-turbo
Whisper large-v3-turbo is a audio and video speech-to-text model (MIT). Served on Mixpeek, scaling to zero when idle.
A Ray Serve app in Mixpeek's inference registry hosts it. It starts at zero replicas and wakes on first use.
New opens the Designer in Studio. For an existing namespace, open it in Studio and use its Designer tab.
No demo for Whisper large-v3-turbo yet
A video, image, slideshow or playground goes here once we record one on our own data.
What it is good and bad at
- Best at
- Multilingual speech-to-text on GPU at large-v3 quality with the decoder cut from 32 to 4 layers, so decoding is much faster than large-v3.
- Worst at
- Word-level timing through the transformers pipeline (Live Nation measured 9.4 GB peak and wrong times); CPU-only clusters (the code forces the whisper-1 API when there is no GPU); music and non-speech audio.
- Task
- Transcription
- Input
- Audio, Video
- Output
- Text
- Use cases
- Transcripts
- Why we recommend it
- The served speech model behind multimodal_extractor transcription.
Templates that use it
No template uses this model yet. Browse templates.
Input and output
Input
- audio
- sample_rate
- 16000
- long_form
- return_timestamps=True (MVS_WHISPER_RETURN_TIMESTAMPS), needed past 30 s
- language
- optional forced language; auto-detect when unset
Output
- text
- transcript per audio chunk
How Mixpeek wires it
- Extractors
multimodal_extractor@v1run_transcription=true and HARDWARE_WHISPER_USE_API=false on a GPU clustermultimodal_extractor@v2run_transcription=true and HARDWARE_WHISPER_USE_API=false on a GPU cluster
- Feature URIs
- None today
- Vector indexes
- None
- Retriever stages
- None
- Combines with
- video -> audio chunks -> whisper transcript -> E5 transcription vector (multimodal_extractor_v*_transcription_embedding)
- customers/livenation/manifest-loop ln_shots plugin: 30 s windows with segment timestamps, then Qwen3-Embedding-0.6B text vectors
How it runs
- Status
- Served on Mixpeek Ray Serve app in the inference registry. multimodal_extractor loads it in-process (WhisperLocalTranscriptionBatch, GPU) only when the whisper-1 API is off; prod runs HARDWARE_WHISPER_USE_API=true per infra/hardware/config.py comments.
- Serve apps
openai__whisper_large_v3_turbo- Runtime
- transformers AutoModelForSpeechSeq2Seq + AutoProcessor + pipeline("automatic-speech-recognition")
- Weights
- safetensors, no remote code
- Parameters
- 809 M
- Size on disk
- 1,618 MB
- Preprocessing
- Decode and resample audio to 16 kHz mono; fp16 on GPU, fp32 otherwise.
- Postprocessing
- Keep result["text"]; timestamps are computed and discarded.
- Serving
- both
- Hardware
- Needs a GPU
Scale to zero
- Minimum replicas
- 0
- Maximum replicas
- 1
- Idle timeout
- 600 s
- Memory
- Not recorded
- CPUs
- 0.5
- GPUs
- Set at deploy time: float(os.getenv('HARDWARE_WHISPER_NUM_GPUS', '0.5')) if hardware_config.gpu_concurrency > 0 else 0
- Cold start, node warm
- Not recorded
- Cold start, new node
- Not recorded
Measured cost
Unmeasured
No committed full measurement. Code comment (BACKE-740): return_timestamps cuts L4 throughput from 2.54 to 1.30 (units not stated).
Our evaluations
Unmeasured
We have not evaluated this model on our own data yet. Publisher results are in the sources below.
Questions
- What is Whisper large-v3-turbo best at?
- Multilingual speech-to-text on GPU at large-v3 quality with the decoder cut from 32 to 4 layers, so decoding is much faster than large-v3.
- What are the limits of Whisper large-v3-turbo?
- Word-level timing through the transformers pipeline (Live Nation measured 9.4 GB peak and wrong times); CPU-only clusters (the code forces the whisper-1 API when there is no GPU); music and non-speech audio.
- Can Whisper large-v3-turbo be used commercially?
- Yes. MIT weights and code; no use restrictions.
- How does Whisper large-v3-turbo run on Mixpeek?
- Ray Serve app in the inference registry. multimodal_extractor loads it in-process (WhisperLocalTranscriptionBatch, GPU) only when the whisper-1 API is off; prod runs HARDWARE_WHISPER_USE_API=true per infra/hardware/config.py comments.