nepali-conformer-streaming
Cache-aware streaming Nepali ASR (520 ms lookahead). Carries a large, honestly-reported streaming-lineage penalty on real calls — read RESULTS.md before choosing this over the offline model; it exists because a phone agent needs incremental output.
Try it: demo Space · Everything else: github.com/Ampixa/nepaliconformer (NepTel benchmark, per-system outputs, full honest results)
Numbers (measured, not marketed)
| benchmark | WER |
|---|---|
| NepTel — real Nepali call audio, human-reviewed refs | 59.87 |
| Held-out gold read Nepali (W1 slice) | 31.5 |
| Whisper-large-v3 zero-shot on the same NepTel audio | 99.4 |
Architecture
121.3M-parameter 17-layer Conformer (d=512, striding ×4, 40 ms frames), hybrid TDT/CTC decoder, 1,024-piece Devanagari SentencePiece. Chunked-limited attention [[70,13],[70,6],[70,1],[70,0]], fully causal convolutions, cache-aware incremental decoding.
Training data
~1,655 h of mostly conversational Nepali (YouTube podcasts/interviews) with Google Chirp 2 pseudo-labels + 105 h human-labeled read speech; telephony codec, noise, reverb and tempo augmentation. Label-noise ceiling and every measured limitation (English, sung speech, slow speech, end-of-turn) are documented in the repo's RESULTS.md.
Usage
from nemo.collections.asr.models import EncDecHybridRNNTCTCBPEModel
m = EncDecHybridRNNTCTCBPEModel.restore_from("nepali_conformer_streaming.nemo")
print(m.transcribe(["audio.wav"])[0].text)
License: CC-BY-NC-4.0 (weights). Code in the repo: MIT.
- Downloads last month
- -
Evaluation results
- Real-call WER on NepTel v0.1 (real Nepali call-center audio, human-reviewed)self-reported59.870
- Real-call CER on NepTel v0.1 (real Nepali call-center audio, human-reviewed)self-reported41.080
- Read-speech WER on Held-out gold read Nepali (W1 read slice, OpenSLR-54 utterances absent from training)self-reported31.500