Speech & audio
Transcribe calls in batches with Whisper
Recordings go on a volume; a dataset build normalises them, cuts them into 10 to 40 second clips, and optionally drafts transcripts with Whisper on a GPU. People then correct the drafts in the annotation workspace, and the result is a TSV any trainer can read.
CPU containers for import and chunking; a GPU container (T4 is enough) for Whisper drafts.
How it works
- 1Import: pick a folder on the datasets volume, or S3, GCS, BigQuery, Hugging Face or a URL list. Audio is normalised and chunked; diarisation splits speakers.
- 2Pre-label: optional. Whisper runs on the cluster's GPU pool and writes a draft transcript per clip; hosted providers (Azure, Google, ElevenLabs) work too.
- 3Annotate: the dataset page shows each clip with its draft; a person fixes the text and the speaker role, with keyboard shortcuts.
- 4Share: named links per annotator, an NDA gate if you need one, and a spreadsheet export and import for teams who work elsewhere.
The code
# the same pipeline the Datasets page runs, from a terminal cd lab/ingest python app.py --name nhc-sept --kind stt --source volume --path uploads/nhc # import + chunk python app.py --name nhc-sept --prelabel --provider whisper --model large-v3-turbo --gpu T4 # drafts
levers9 cp ./recordings levers9://datasets/uploads/nhc # drop the audio on the volume # then Datasets → New dataset → kind stt → source volume
Where to go next
Related
Score a text-to-speech provider on your data
Synthesize a dataset with several providers and let reviewers rate naturalness and accent, side by side.
Teach a call agent with human feedback
Play the seven missions: calls in, hot spots, duels for 10 XP a pick, level up, boss fight, crown the champion, connect it to real customers.
Share files between jobs on a volume
Persistent folders mounted into every container; browse, upload and download from the dashboard or the CLI.