Whisper
An open speech-to-text model: what it does on office hardware, and where it fails.

OpenAI’s README gives these figures. Relative speed was measured on an A100 transcribing English and varies with language, speaker and hardware.

Whisper is an encoder-decoder transformer. Audio is resampled to 16,000 Hz and turned into a log-Mel spectrogram (80 channels, or 128 in large-v3). One model handles transcription, translation to English, language identification and voice activity detection, steered by special tokens at the start of its output. It reads 30-second windows, so for long audio it shifts the window using the timestamps it predicts.

OpenAI’s paper says that long-form transcription needed extra heuristics: beam search with 5 beams, raising the temperature when the average log probability falls below −1 or the text compresses too well (a sign of repetition), and a no-speech threshold of 0.6. It calls these a workaround for noisy predictions. The turbo model is large-v3 with the decoder cut from 32 layers to 4: faster, with a minor quality loss, and not trained for translation.

The 2022 paper lists the hours of speech-recognition training data for each language and reports word error rates for the large-v2 model on FLEURS, a read-speech test of about 12 hours per language, not meetings. The paper says that performance in a language is very well predicted by its amount of training data.

The faster-whisper project, a faster reimplementation, publishes a benchmark: 13 minutes of audio with large-v2 on an NVIDIA RTX 3070 Ti with 8 GB. OpenAI’s own code took 2 minutes 23 seconds and 4,708 MB of VRAM. faster-whisper took 1 minute 3 seconds and 4,525 MB at 16-bit, 59 seconds and 2,926 MB at 8-bit, and 17 seconds with a batch size of 8. Scaled to a one-hour meeting, the 16-bit run is roughly five minutes; that is an estimate from the benchmark, not a measurement. The same project’s small model took 1 minute 42 seconds at 8-bit on an Intel Core i7-12700K processor with no GPU.

faster-whisper can integrate the Silero VAD model to skip silence, and the project points to WhisperX for speaker diarization and word-level timestamps. These benchmarks use a gaming card and a desktop CPU, not any ANVE hardware.

• Consent for recording, and where the audio and transcripts are kept.
• Your language mix. Neither the paper nor the model cards report a test on code-switched speech.
• Hallucination: the model card warns that predictions may include text not spoken in the audio, and that the model is prone to repeating itself.
• Names, numbers and dates, which a person should check.
• The licence of the exact checkpoint: the README says MIT for code and weights, the large-v3 card lists Apache 2.0, and the turbo card lists MIT.
• Which model you need: turbo cannot translate; use medium or large for that.

ANVE’s minutes maker produces transcripts in Hindi, English and Hinglish on a device in the office. Which speech model and settings it uses is a detail to get in writing before a paid pilot, and ANVE has published no speed figures. Our open-model pages cover Whisper large-v3 and large-v3-turbo.

No sealed ANVE unit exists; every image of the appliance on this site is a labelled render. The software runs today on a development machine. No customers yet, no certifications, no delivery date, and the final price is not announced.

Reservations are ₹999 in India or US$12 elsewhere, refundable in full until you confirm an order in writing (terms). Offices that cannot use cloud AI can apply to be one of ten design partners.