Nemotron 3.5 ASR Streaming — Hausa (ha-NG) Fine-Tune, Round 2
Continued fine-tune of EYEDOL/nemotron-3.5-asr-streaming-hausa
(itself a fine-tune of nvidia/nemotron-3.5-asr-streaming-0.6b) for 15
additional epochs on Hausa speech data.
Training data
- 41.2 hours of Hausa speech from [EYEDOL/naija-voices-hausa-split_1-6 + EYEDOL/naija-voices-hausa-split_1-7](https://huggingface.co/datasets/EYEDOL/naija-voices-hausa-split_1-6 + EYEDOL/naija-voices-hausa-split_1-7).
- Held out 56 speakers (speaker-disjoint, audited for zero train/test overlap — see Section 3b) for evaluation.
- Transcripts normalized: Hausa hooked consonants (
ɗ ɓ ƙ ƴand uppercase variants) mapped to plain Latin equivalents (d b k y) before training, same as round 1.
Evaluation
Metrics measured on the held-out speaker-disjoint test set using true cache-aware
streaming inference (chunked audio, carried encoder/decoder state — not an offline
full-utterance shortcut), att_context_size=[56,0] (80ms chunk, 0ms lookahead).
No-space CER (CER computed after stripping spaces from both reference and hypothesis) is reported alongside WER to separate genuine character-level recognition errors from word-boundary/segmentation differences.
| Model | WER (%) | CER (%) | No-space CER (%) |
|---|---|---|---|
| Round 1 (before this run) | 35.56 | 10.64 | 10.40 |
| Round 2 (15 more epochs, this checkpoint) | 34.99 | 10.35 | 10.12 |
Full evaluation reports (model/manifest paths, hashes, elapsed time, RTFx) are written to
eval_round1/report.json and eval_round2/report.json and included in training_log.txt
above.
Usage
import nemo.collections.asr as nemo_asr
model = nemo_asr.models.ASRModel.from_pretrained("EYEDOL/nemotron-3.5-asr-streaming-hausa")
# Streaming inference: see speech_to_text_cache_aware_streaming_infer.py in NeMo's repo.