nhtera/ghira-asr-vi-0.6b

Ghira ASR vi (Nemotron 3.5 streaming 0.6B, Vietnamese fine-tune)

A Vietnamese fine-tune of nvidia/nemotron-3.5-asr-streaming-0.6b (revision ea30d66d). It is the speech model of Ghira, an offline meeting note taker. It is published as a q8_0 GGUF for NeMo-Speech.cpp, with the same architecture, tokenizer and language prompts as the base. It is a drop-in replacement.

This is an independent fine-tune. It is not made or endorsed by NVIDIA.

Results

Word error rate in %, lower is better. "Syllable" WER is used for Vietnamese and mixed error rate for ViMedCSS. "Long" means 3–6 min files of joined clips, read with a stream reset after each Vietnamese line. "Live" means Ghira's live transcript pipeline in real time.

Set	Base	This model
FLEURS-vi, long	12.2	11.1
ViMedCSS (medical, code-switch), long	18.6	16.8
VietMed (medical conversation), long	14.7	13.3
FLEURS-vi, live	17.7	15.5
ViMedCSS, live	21.8	20.0
VietMed, live	16.7	16.1
FLEURS-en, per clip	7.4	6.5
Earnings-21, per clip	10.3	10.3
FLEURS-en, live	10.6	11.3
FLEURS-en, long	7.6	6.9

The test splits were never trained on. ViMedCSS training videos that also appear in its test split were dropped, and so was every unlabeled VietMed recording with a clip that repeats a VietMed test sentence.

Training

• Recipe: the mean of two fine-tunes of the same base with NeMo speech_to_text_finetune.py, lr 3e-5, on one A100: the previous release (5,000 steps), and a 7,500-step run whose weights are the mean of steps 6,000, 7,000 and the last.
• Targets: the base model's own transcripts (with its language tags). Where they differ from the human caption mid-sentence, the caption's words are used. Half the examples join 2–3 clips with short pauses, half use the auto language prompt, and 15% of single clips pass through a telephone band.
• Unlabeled VietMed: transcribed by an earlier version of this model and kept only where Whisper large-v3-turbo (MIT) reads nearly the same words (15% of clips).
• Data of the 7,500-step run (ViMedCSS and VietMed are repeated 3×):

Data	Hours used	Licence
YODAS vi000 (Vietnamese)	268	CC BY 3.0
YODAS en000 (English replay)	136	CC BY 3.0
LibriSpeech (English replay)	50	CC BY 4.0
VietMed unlabeled (8 kHz medical talk)	29	CC BY 4.0
FPT Open Speech Dataset	28	CC BY 4.0
ViMedCSS train, validation, hard	22	CC BY 4.0
VietMed train, dev	10	MIT

YODAS audio comes from YouTube videos that their uploaders licensed under Creative Commons BY. The attribution goes to those uploaders, via the YODAS dataset (Li et al., 2023, "YODAS: Youtube-Oriented Dataset for Audio and Speech"). VietMed, labelled and unlabeled, is by Le-Duc (2024, "VietMed: A Dataset and Benchmark for Automatic Speech Recognition of Vietnamese in the Medical Domain").

The data preparation and training scripts are in github.com/nhtera/Ghira.

Limitations

• Medical speech read after the meeting improved by about 10% relative (live: 4–8%). That is real, but still well above general-speech error rates.
• Live English is 0.7 points worse than the base on FLEURS-en; English read after the meeting is better.
• The model was tuned and checked for Vietnamese and English only. Other languages of the base were not measured and may be worse.
• Like the base, it can drop or mishear words in overlapping speech and in noisy far-field audio.

Licence

Governed by the OpenMDW License Agreement, version 1.1, the base model's licence. Derived from NVIDIA's nemotron-3.5-asr-streaming-0.6b, which is governed by OpenMDW-1.1.