Whisper large-v3
OpenAI · Released Nov 2023
Whisper large-v3 is OpenAI's open-weight, Apache 2.0 speech recognition model (1.55B parameters, 99 languages), which OpenAI says cuts errors 10 to 20% against large-v2. Run through faster-whisper in float16 on CUDA, it worked in our runs as a reliable per-line judge of sung vocals on a 6 GB GPU. Recommended for local transcription and checking, with two caveats: fast rap needs short windows, and it can misread words as profanity, so listen to flagged lines.
Is it right for you?
Good for
- Local, free transcription in 99 languages under the Apache 2.0 license, with no per-minute bill
- Checking generated vocals line by line against a lyric sheet: in our runs it scored every take per line, and the best-scoring take sang each line of a written jingle right
- Faster runs through faster-whisper, which its authors say is up to 4 times faster than the reference implementation at the same accuracy
- A 6 GB NVIDIA GPU: float16 on CUDA ran in our tests on an RTX 3060 Laptop GPU
Not good for
- Long, fast rap transcribed in one pass: it dropped the opening seconds in our runs; transcribe 5-second windows snapped to silence instead
- Publishing on its word alone: it misheard clean lyrics as profanity in our runs, so listen to every flagged line
- Speaker labels or live streaming captions out of the box; it is a batch transcriber without diarization
Pricing
Input
Free (self-hosted open weights)
Output
Free (self-hosted open weights)
Context
Runs locally via faster-whisper, 6 GB GPU tested; Apache 2.0
No verdict changes yet
The clock starts day one. Changes land here as our verdict evolves.
Sources
Verification log
No verification checks yet
We haven't logged a verification check for this entry. Once a check runs, its history shows here.