ADR 0178 — Whisper large-v3 is the default transcription model

• Status: Accepted
• Date: 2026-11-11
• Amends: ADR 0096 — the default model is large-v3; Large v3 Turbo becomes a choice beside Small
• Related: ADR 0095 (the transcription pipeline), ADR 0131 (recordings), audio ingestion

Context

Meetings recorded in Imprint are mostly in Romanian. Read after the first live meeting, the Large v3 Turbo transcript had mistakes in names and words. The user asked for a more accurate model, even if slower, as the default, with the other two still offered.

Whisper Turbo keeps large-v3’s encoder but cuts the decoder from 32 layers to 4. OpenAI’s language breakdown for Turbo (openai/whisper discussion 2363) gives these word error rates for Romanian:

	Common Voice 15	FLEURS
large-v3	10.8	8.2
Turbo	16.1	9.5

That is about a third fewer mistakes on Common Voice, whose varied microphones are the closest of these to meetings, and about 14% fewer on FLEURS. English differs little (9.3 against 10.6, and 4.1 against 4.4).

Decision

Add large-v3 to audio.model, pinned like the other models in audio_ingest.py: Systran/faster-whisper-large-v3 at revision edaa852ec7e145841d8ffdb056a99866b5f0a478 (MIT, ungated, the conversion faster-whisper’s own large-v3 name points to), with the size and SHA-256 of model.bin, config.json, preprocessor_config.json, tokenizer.json and vocabulary.json. It is the default. Large v3 Turbo and Small stay as choices. Inference is unchanged: CPU, int8, four threads, beam 5, voice activity detection, word timestamps.

Rejected alternatives

• Keeping Turbo as the default. It is the faster choice and stays available, but the user asked for accuracy first.
• large-v2. Reading OpenAI’s chart, its Romanian error rates are about 13.7 (Common Voice) and 12.4 (FLEURS), worse than large-v3’s.
• Distilled large-v3 models. English only.
• Romanian fine-tunes. TransferRapid’s Turbo fine-tune reports a large Romanian gain, but it is CC-BY-NC-4.0, was selected on its test set and is not converted to CTranslate2. BuzzASR’s large-v3 fine-tune scores about the same as plain large-v3 on FLEURS. Others found have no license. Training on one language also risks English.

Consequences

• Recordings and audio files are transcribed with fewer mistakes in Romanian.
• Transcription takes about twice as long. Measured on this project’s Windows machine (Ryzen AI Max+ 395, int8, four threads): 185 s for 180 s of Romanian read speech against 92 s with Turbo, and 33 s against 14 s for 150 s of a real call’s audio. A one-hour meeting, both channels, takes on the order of an hour.
• The first transcription downloads about 2.88 GiB, against about 1.51 GiB for Turbo; memory use peaks around 3 GB.
• Installations that never chose a model switch to large-v3, so their audio files are transcribed again the next time they are indexed. Recordings keep their transcript until the user chooses Transcribe again.
• On a 150-second excerpt of the real call, the two models’ text was close, so that excerpt neither confirms nor refutes the published gain. The gain rests on OpenAI’s figures.

Invalidation conditions

• If a GPU runtime for faster-whisper becomes available on users’ machines, reconsider the speed trade-off and the int8 choice.
• If measured on the user’s meetings large-v3 is not more accurate than Turbo, make Turbo the default again.
• If a permissively licensed Romanian model in CTranslate2 format proves better on held-out data, offer it.

Validation

• tests/test_audio_ingestion.py covers the settings and profile; the full suite passes on Windows 11 with Python 3.13.
• The pinned files were downloaded and their SHA-256 recomputed. Both models ran on 180 s of Romanian read speech and on 150 s of the real call’s computer-audio channel, with the pipeline’s parameters.
• Not run: a word-error-rate measurement on the user’s own meetings, for want of reference transcripts.