Nemotron 3.5 ASR Streaming — Yoruba (yo-NG) Fine-Tune, Round 2
Continued fine-tune of EYEDOL/nemotron-3.5-asr-streaming-yoruba
(itself a fine-tune of nvidia/nemotron-3.5-asr-streaming-0.6b) for 15
additional epochs on Yoruba speech data.
Training data
- 39.8 hours of Yoruba speech from: EYEDOL/naija-voices-yoruba-split_2-8, EYEDOL/naija-voices-yoruba-split_2-9.
- Held out 52 speakers (speaker-disjoint, audited for zero train/test overlap) for evaluation.
- Transcript normalization mode:
strip_tones(same as earlier rounds).
Evaluation
Held-out speaker-disjoint test set, true cache-aware streaming inference, att_context_size=[56,0].
No-space CER strips spaces from reference and hypothesis to separate character-level errors from word-boundary differences.
| Model | WER (%) | CER (%) | No-space CER (%) |
|---|---|---|---|
| Round 1 (before this run) | 43.65 | 14.91 | 14.66 |
| Round 2 (this checkpoint) | 43.65 | 14.81 | 14.58 |
Both rows were measured on the same held-out set for this run. Evaluation reports and the full run log are in training_log.txt.
Usage
import nemo.collections.asr as nemo_asr
model = nemo_asr.models.ASRModel.from_pretrained("EYEDOL/nemotron-3.5-asr-streaming-yoruba")
# Streaming inference: see speech_to_text_cache_aware_streaming_infer.py in NeMo's repo
# (pass target_lang=yo-NG).