smajji/nemotron-hinglish-v5

Nemotron Hinglish v5

Nemotron-Hinglish-v5 is a fine-tune of nvidia/nemotron-3.5-asr-streaming-0.6b specialized for English, Hindi and Hinglish (Hindi-English code-switching), built for cache-aware streaming ASR, with improved number / phone-number / digit accuracy.

It continues from the v4 checkpoint and adds a substantial Indian-accent English corpus, which fixes the Devanagari latch: the previous release transcribed Indian-accent English into Hindi/Devanagari script (e.g. "you navigate me to the gas station" → "यू नेविगेट मे टू द गैस स्टेशन"). v5 outputs Latin-script English for Indian-accent speech.

It preserves the base model's FastConformer-Transducer (RNNT) cache-aware streaming architecture (24 layers, 1024 hidden, 600M params) and its multi-lingual 13088 BPE vocab + 128 language prompts, including the auto language-detection prompt.

Model Details

• Base model: nvidia/nemotron-3.5-asr-streaming-0.6b
• Architecture: FastConformer-Transducer (RNNT), cache-aware streaming, 8x subsampling
• Parameters: ~600M
• Sampling rate: 16 kHz mono
• Target languages: English (en), Hindi (hi), Hinglish code-switched (auto prompt)
• Streaming: cache-aware, chunk sizes 80/160/320/560/1120 ms

Training Data

Fine-tuned on 13 corpora totaling ~5,030 hours / 2,227,473 utterances:

Corpus	Utterances	Hours	Language
hinglish_cc	802,866	1,832.2	Hinglish (code-mixed)
shrutilipi_hi	504,030	1,000.0	Hindi
peoples_speech	266,646	1,000.0	English
hindi_1482hrs	253,373	400.0	Hindi
spgispeech	117,927	300.0	English
indic_tti_eng	113,702	172.8	Indian-accent English
earnings22	51,578	103.2	English
iisc_spicor	49,186	96.9	English
hinglish_ujs	52,812	89.4	Hinglish
asr_task_en	8,000	13.9	Indian-accent English
fleurs_en_us	2,601	7.5	English
indicvoices_r_hindi	2,632	7.1	Hindi
fleurs_hi_in	2,120	6.7	Hindi
Total	2,227,473	~5,030

Split across English (~1,690 h, including Indian-accent English — the Devanagari-latch fix), Hindi (~1,414 h) and Hinglish code-mixed (~1,922 h), with number/symbol-dense corpora added for digit & phone-number accuracy (SPGISpeech, IISc_SPICOR, Peoples Speech, earnings22). Trained with auto language self-detection for code-switching, punctuation/casing preserved.

Benchmark

Greedy decode, punctuation-insensitive WER (held-out sample, auto prompt):

Language	Nemotron-Hinglish-v4	Nemotron-Hinglish-v5
English	4.0%	4.2%
Hindi	12.0%	12.5%
Hinglish	29.2%	28.6%

Indian-accent English (200 held-out utts, the Devanagari-latch detector):

Metric	Nemotron-Hinglish-v4	Nemotron-Hinglish-v5
Devanagari latch (% Indian-EN transcribed in Devanagari)	19.5%	0.0%
WER	22.2%	8.3%

• Devanagari latch fixed: v5 outputs Latin-script English for Indian-accent speech (0.0% vs 19.5%).
• Digit / phone-number accuracy improved from added number-dense corpora (TEDLIUM, earnings22, FLEURS).

Usage (NeMo)

import nemo.collections.asr as nemo_asr
model = nemo_asr.models.ASRModel.restore_from("smajji/nemotron-hinglish-v5")
model.eval()
transcriptions = model.transcribe(["audio.wav"], batch_size=4)
print(transcriptions)