whisper-small-kn-curated-reverse-mft-1-1-1

Kannada 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: Kannada (kn)
• 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)	15.94 %	18.23 %
LER (legal entities)	–	0.48 %
NER (numeral)	0.45 %	2.03 %
PER (punctuation)	5.52 %	8.70 %
TER	21.91 %	29.44 %
Sandhi resolutions	404	48

FLEURS-RO LER (0.01%) 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-kn-curated-reverse-mft-1-1-1",
    generate_kwargs={"language": "kn", "task": "transcribe"},
)
print(asr("sample.wav")["text"])

CTranslate2 build: adalat-ai/whisper-small-kn-curated-reverse-mft-1-1-1-ct2-fp16.

Intended use

Kannada dictation in legal, medical, and classroom settings where rich-transcription output is required.

License

Apache-2.0. Base model openai/whisper-small is MIT.