Shenava Koochik v1.5
Persian (Farsi) ASR — FastConformer hybrid RNNT + CTC (114M). The both-heads successor to v1.0.
What is new
- RNNT head revived. v1.0's RNNT head was broken (a tokenizer/blank misalignment made greedy decoding loop forever, ~4398% WER). v1.5 surgically re-initializes the RNNT prednet+joint under the
ve_tok_v4tokenizer and trains it up via a 2-phase curriculum -> a working transducer head. - CTC head preserved. The deployed, record-setting CTC head is kept identical to v1.0 (encoder frozen during the corrective RNNT finetune) -> zero CTC regression.
Benchmarks (golden-6669, strict fair_text normalizer, att_context [70,13])
| head | v1.0 | v1.5 |
|---|---|---|
| CTC | 8.12% | 8.12% (identical) |
| RNNT | 4398% (broken) | 9.50% |
CTC remains the recommended/deployed head; RNNT is newly functional. Character error rate ~2.6%.
Usage
from nemo.collections.asr.models import ASRModel
m = ASRModel.from_pretrained("Reza2kn/Shenava-Koochik-v1.5")
m.change_decoding_strategy(decoder_type="ctc") # or "rnnt"
print(m.transcribe(["audio.wav"]))
Notes
- RNNT output leans colloquial (trained on audio-faithful ASR labels).
- Trained on a cleaned mix (pseudo corpus + synthetic hardwords + human/article gold) with benchmark de-duplication and label-quality filtering (over-extension / repetition-loop removal).