01
Transcribe a media archive.
Queue recordings with stable asset identifiers and save the transcript beside each original file. Build a retry policy that resumes failed work without duplicating completed records.
Whisper Large v3 Turbo
Transcribe speech on GPUs.
Whisper Large v3 Turbo is OpenAI’s speech-to-text model derived from Whisper large-v3. Run your own AI Fabric transcription worker to control audio preparation, timestamped outputs and batch processing.
€5 in free credits for every new customer.
Plan your deploymentAI Fabric access is opening gradually. GPU configuration is confirmed during onboarding.
Transcripts, subtitles and voice notes.
Make recordings easier to read, search and edit. Choose the output your product needs, from plain text to timestamped segments for a caption editor.
02
Prepare editable captions.
Request timestamped segments, then let an editor correct wording and timing. Apply subtitle formatting rules as a separate step instead of treating raw transcript chunks as a finished subtitle file.
03
Add voice notes to a product.
Convert uploaded notes into text users can search and amend. Keep the recording linked to the transcript, especially for names, numbers and low-confidence passages.
Model capabilities and upstream setup: Whisper Large v3 Turbo publisher reference. Reviewed .
Deployment preparation
Run Whisper transcription with timestamps in Python.
The example loads a Transformers transcription pipeline and requests timestamps from a local WAV file. Package audio decoding dependencies and define the file sizes and durations your own service will accept.
- Install compatible Transformers, PyTorch and audio decoding dependencies; confirm that the worker can decode your input format.
- Test short, clean recordings before evaluating longer files, silence, background noise and different languages.
- Choose chunking and batch settings from measured memory and transcription quality, then persist text and timing metadata together.
import torch
from transformers import pipeline
transcribe = pipeline(
"automatic-speech-recognition",
model="openai/whisper-large-v3-turbo",
torch_dtype=torch.float16,
device="cuda:0",
)
result = transcribe(
"recording.wav",
return_timestamps=True,
generate_kwargs={"task": "transcribe"},
)
print(result["text"])
for segment in result["chunks"]:
print(segment["timestamp"], segment["text"])Provide recording.wav and validate the audio decoder and dependency versions. This runs your loaded model, not the managed audio API.
GPU memory & performanceThe model card describes a reduced decoder, from 32 layers to 4, trading some quality for faster inference. Actual memory and speed depend on audio batching, decoding and runtime; measure the complete job on your card.
In your applicationCheck for invented words in silence and misheard names in difficult audio. A transcript review workflow should allow corrections while retaining the original recording.
AI Fabric availability & costRTX A4500 (20 GB) and RTX 2000 Ada (16 GB) are the GPU options. Access, runtime settings and billing are confirmed during onboarding. No Blade latency or throughput benchmark is published for this configuration.
Combine Whisper transcripts with speaker diarization.
Whisper produces words; diarization produces speaker intervals. If your product needs speaker-labelled transcripts, run pyannote Community-1 as a separate stage and reconcile the two timelines. Store uncertain alignments for review.
01 / InputRecording
02 / ProcessTranscript + time segments
03 / ResultOptional speaker alignment
Whisper Large v3 Turbo
deployment FAQ.
Does Whisper Large v3 Turbo identify speakers?
Not by itself. Pair transcription with a diarization model and align the outputs if your application needs speaker labels.
Can I produce subtitles?
The pipeline can return timestamps, but subtitle segmentation, line length, timing review and export formatting remain application work.
Do managed API upload limits apply to my worker?
No. In a custom deployment, your request handler, audio decoder, memory and storage design define the allowed inputs. Set explicit limits before accepting user uploads.
How much does running Whisper Large v3 Turbo on AI Fabric cost?
Measure GPU processing time per audio hour using representative recordings. Include decoding, chunking and retries; add diarization as a separate stage when speaker labels are needed.
The GPU billing unit and idle policy are confirmed during onboarding. Managed Models API rates do not apply to this deployment.
Checkpoint: openai/whisper-large-v3-turbo. Access requirements and licence: Whisper Large v3 Turbo publisher reference.