sugar-smoke

nvidia/nemotron-3.5-asr-streaming-0.6b fine-tuned for Akan (Twi), Ga and Ewe, with replay of English, French, Spanish, German and Portuguese so the original languages are kept. New language prompts: ak-GH, gaa-GH, ee-GH. All original prompts (and auto) still work.

The sugar-smoke.nemo file at the root is the best checkpoint; final/sugar-smoke.nemo is the model at the last training step. This card describes the best one: step 20 of run d401e5fd72e4 - selection score 107.584 (target WER + 2.0 x average replay WER increase over the base model).

Language	Prompt		WER base	WER this model
Akan (Twi)	ak-GH	new	101.0	100.0
Ga	gaa-GH	new	98.36	100.0
Ewe	ee-GH	new	103.33	100.0
English	en-US	kept	6.62	6.62
French	fr-FR	kept	6.9	7.76
Spanish	es-ES	kept	3.19	4.26
German	de-DE	kept	20.78	35.06
Portuguese	pt-BR	kept	4.13	6.88

WER is on held-out clips, lower-cased, punctuation and language tags removed. "auto" language-ID accuracy on the new languages: 15.7 %.

Use (NeMo)

import nemo.collections.asr as nemo_asr
from huggingface_hub import hf_hub_download

model = nemo_asr.models.ASRModel.restore_from(hf_hub_download("prince4332/sugar-smoke", "sugar-smoke.nemo"))
print(model.transcribe(["clip.wav"], target_lang="ak-GH")[0].text)   # or "gaa-GH", "ee-GH", "en-US", "auto"

Streaming: model.set_inference_prompt("ee-GH"), then the usual cache-aware loop (CacheAwareStreamingAudioBuffer + conformer_stream_step). Language tags like <ee-GH> are stripped by the decoder (decoding.strip_lang_tags).