tigopt/whisper-lv3-ft-hush-ct2

Whisper large-v3 pt-PT for the noise-suppressed telephone line (CTranslate2)

CTranslate2 conversion of whisper-lv3-ft-hush: inesc-id/WhisperLv3-FT with a LoRA, merged, trained on telephone-line audio that went through the Hush noise suppressor at 22.5 dB, under decoder prompts from its training text rather than any client's. Training data, code, settings, evaluation and limits are in that repository. Git tag v1 is the first run; main is v2.

Where the base model reads its bias prompt back out of background voices, this one does not: invented words in a meeting room fall from about 800–1,100 to about 160–200 per 55 caller phrases on the raw line, and to 5–80 with Hush in front of it. Clean WER is level with the base. Run it on the raw line, and give it the Hush-cleaned line once background voices are confirmed on a call: Hush costs words on a quiet line.

Weights are float16; choose the compute type at load (int8_float16 on GPU).

Use

faster-whisper:

from faster_whisper import WhisperModel

model = WhisperModel("<this repo>", device="cuda", compute_type="int8_float16")
segments, _ = model.transcribe("call.wav", language="pt")
print(" ".join(s.text.strip() for s in segments))

speaches lists and serves it by this card's library_name: ctranslate2 and pipeline_tag: automatic-speech-recognition:

curl -X POST localhost:8000/v1/models/<this repo>
curl localhost:8000/v1/audio/transcriptions -F file=@call.wav -F model=<this repo> -F language=pt

Conversion

ct2-transformers-converter --model whisper-lv3-ft-hush --output_dir . \
  --copy_files tokenizer.json preprocessor_config.json --quantization float16

preprocessor_config.json carries feature_size: 128, which faster-whisper needs for large-v3.

Attribution

FalAR (INESC-ID) and DEMAND are CC BY 4.0; inesc-id/WhisperLv3-FT and Whisper (OpenAI) are Apache-2.0.