Whisper Small.en — Fine-tuned (CTranslate2 int8)
CTranslate2 int8 quantized version of ronxldwilson/whisper-small-en-finetuned for use with faster-whisper.
Drop-in replacement for small.en — same speed, same RAM, better accuracy (5.76% WER vs ~24% baseline on FLEURS en_us).
Usage
from faster_whisper import WhisperModel
model = WhisperModel(
"ronxldwilson/whisper-small-en-finetuned-ct2-int8",
device="cpu",
compute_type="int8",
)
segments, _ = model.transcribe(audio, beam_size=1)
text = " ".join(s.text for s in segments).strip()
Benchmark
| Baseline (small.en) | This model | |
|---|---|---|
| WER (FLEURS en_us) | ~24% | 5.76% |
| Exact match (15 phrases) | 47% | 80% |
| Latency (CPU int8) | 1495ms | 1407ms |
| Model size | 244 MB | 243 MB |
Training
See the base model card for full training details: ronxldwilson/whisper-small-en-finetuned