whisper-small-hi-curated-reverse-mft-1-1-1

Hindi rich-transcription ASR — emits text with punctuation and numerals natively. Released alongside SCRIBE (Interspeech 2026, under review).

Model details

Architecture: Whisper-small (244M, encoder-decoder)
Base model: openai/whisper-small
Language: Hindi (hi)
Output style: rich transcription (punctuation + formatted numerals)
Training: three-stage curriculum fine-tune on LLM-curated rich-transcription data (diversity → pace/style → precision); this is the stage-3 checkpoint.

Evaluation

Evaluated with SCRIBE. Metrics: WER (lexical, sandhi-aware), LER (legal entities), NER (numerals), PER (punctuation), TER = sum of categorical rates.

Metric	FLEURS-RO	IN22-Legal
WER (lexical)	12.05 %	9.51 %
LER (legal entities)	–	0.55 %
NER (numeral)	0.36 %	1.05 %
PER (punctuation)	3.49 %	6.39 %
TER	15.90 %	17.50 %
Sandhi resolutions	125	34

FLEURS-RO LER (0.05%) is folded into WER — too sparse on a general-domain set to warrant a row. SCRIBE WER ≠ jiwer WER: SCRIBE's WER is the lexical-category rate after sandhi-tolerant alignment, not monolithic edit distance. Paper uses ERlex / ERnum / ERpunc / ERent; this library uses WER / NER / PER / LER for the same quantities. Paper Table 1 reports the whisper-medium counterpart of this checkpoint.

Usage

from transformers import pipeline
asr = pipeline(
    "automatic-speech-recognition",
    model="adalat-ai/whisper-small-hi-curated-reverse-mft-1-1-1",
    generate_kwargs={"language": "hi", "task": "transcribe"},
)
print(asr("sample.wav")["text"])