provsalt/nemotron-3.5-singlish-multilingual

Nemotron 3.5 Singlish multilingual

Experimental Singapore English / Singlish adaptation of NVIDIA Nemotron 3.5 ASR Streaming 0.6B. The original multilingual tokenizer and language prompts are retained. A subsequent 200-clip-per-language audit found a Japanese regression in this checkpoint. Mandarin was roughly unchanged on the read-speech audit. Other languages have only limited or no regression checks.

Files

• nemotron-3.5-singlish-multilingual.q8_0.gguf: Q8 model for NeMo-Speech.cpp.
• nemotron-3.5-singlish-multilingual.nemo: self-contained NeMo checkpoint.
• evaluation.json, training-settings.json, and provenance.json: measured scores, training settings, and artifact provenance.
• SHA256SUMS: checksums for the release files.

Validation

The native Q8 GGUF was evaluated with nemo-speech 0.1.0, 160 ms streaming, context [56,1], explicit language prompts, and verbatim output. Edit counts were independently verified with jiwer.

Set	Metric	Error rate	Clips
Singapore validation	WER	22.08%	205
FLEURS English	WER	9.34%	20
FLEURS Mandarin	CER	17.09%	20
FLEURS French	WER	14.29%	20

Singapore validation uses speaker/partner components excluded from adaptation training. Validation was used to select this checkpoint, so these are development results. The fresh two-hour test set has not been evaluated. Base-model pretraining overlap is unknown. This release has not reached the 20% validation WER target, and accuracy can differ on other microphones, speakers, or conversational settings. Rare Singlish and Hokkien expressions remain a limitation.

Expanded multilingual audit

A September 12, 2026 audit compared this release with the original NeMo checkpoint converted using the same Q8 converter. Both used the same 160 ms streaming runtime and 200 FLEURS test clips per language. Character edit counts were independently verified with jiwer.

Language mode	Language	Original CER	This release CER
conditioned	zh-CN	19.58%	19.32%
conditioned	ja-JP	15.15%	19.93%
auto	zh-CN	20.02%	19.76%
auto	ja-JP	16.28%	21.91%

Lower CER is better. Japanese was absent from prior training replay and regression checks. Some Japanese failures omit most or all of a sentence. These diagnostic results do not establish conversational or code-switching accuracy. FLEURS speaker metadata was not retained, and base pretraining overlap is unknown.

The model weights in this repository are still the checkpoint shown above. A separate multilingual recovery experiment is in progress. See multilingual-audit.json for measured differences and paired clip bootstrap intervals.

Training

The latest adaptation uses 60 hours of Singapore conversation and 5.46 hours of English, Mandarin, and French replay. It trains the last 12 encoder layers plus decoder and language-conditioning weights. The deployed [56,1] context is sampled 60% of the time, with the original contexts retained for the remaining 40%. See training-settings.json for the checkpoint's exact settings. Earlier adaptations used general Singapore speech and a slang-focused mix.

Singapore audio came from the NSC Part 3 separate-room material mirrored in mesolitica/IMDA-STT. Multilingual replay used official FLEURS training splits; its regression checks used test clips. Training audio and transcript manifests are not included in this model repository.

Run the GGUF

With a compatible NVIDIA NeMo-Speech.cpp runtime installed:

hf download provsalt/nemotron-3.5-singlish-multilingual \
  nemotron-3.5-singlish-multilingual.q8_0.gguf --local-dir ./model
nemo-speech transcribe ./audio \
  --model ./model/nemotron-3.5-singlish-multilingual.q8_0.gguf \
  --device cuda --language en-US --stream --verbatim --format json \
  --asr.streaming.rnnt_right_context 1 --output-dir ./transcripts

Authenticate with hf auth login to access this private repository. Use zh-CN or fr-FR for the corresponding language prompts.

Attribution and license

The base model was released by NVIDIA. Its governing license is OpenMDW-1.1, as stated in the upstream model card. This repository contains a modified model and records the upstream revision in provenance.json.