YAML Metadata Warning:empty or missing yaml metadata in repo card

Check out the documentation for more information.

Nemotron-Hinglish-v12 (Mevine ASR — En + Hi + Hinglish)

A bilingual Hindi/English Hinglish streaming ASR model, fine-tuned from nvidia/nemotron-3.5-asr-streaming-0.6b (cache-aware FastConformer-RNNT, 0.6B).

This is the flagship (12th) iteration. It is warm-started from the best v11 checkpoint and tuned for strong code-mixed Hindi-English performance while keeping clean English / Hindi strong.

Full benchmark report (all Mevine versions v1→v12 across 12 datasets): see BENCHMARKS.md.

Training data (2,905,937 utts / 6,451 h)

Corpus / dataset utts hours lang
Hinglish code-mixed (hinglish_cc) 802,866 1832.2 hinglish
Shrutilipi (Hindi) 428,446 849.5 hi
Indian-English NPTEL 348,217 744.7 en
Hindi-1482hrs 253,373 400.0 hi
People's Speech 213,435 800.6 en
TEDLIUM 119,216 259.9 en
SPGISpeech 117,927 300.0 en
IndicTTS-English 113,702 172.8 en
India-Accent Common Voice 91,510 144.5 en
OpenSLR104 (code-mixed, NEW) 52,825 89.6 hinglish
Hinglish UJS 52,812 89.4 hinglish
OpenSLR Hinglish 52,812 89.4 hinglish
MUCS Hinglish 52,809 97.4 hinglish
earnings22 51,578 103.2 en
IISc SPICOR 49,186 96.9 en
Hinglish Casual 33,275 100.7 hinglish
NPTEL Tech English 27,259 50.5 en
Hindi Audio Stories 25,314 188.5 hi
ASR Task EN 8,000 13.9 en
Roopa code-mixed numbers 4,022 6.6 hinglish
IndicVoices-R Hindi 2,632 7.1 hi
FLEURS (en_us / hi_in) 2,601 / 2,120 7.5 / 6.7 en/hi
  • Devanagari digits + spelled numbers normalized to ASCII digits for cleaner, consistent targets.
  • Code-mixed share ~36% (raised without cutting English), plus fresh OpenSLR104 code-mixed data.
  • tarred audio (16 kHz FLAC), Lhotse dynamic bucketing, batch_duration=300, lr 8e-5 (Cosine→1e-6, 5% warmup), max_steps=30000.

Evaluation (authoritative 8-type battery, 300/type, median WER)

type v10 v11 v12
code-mixed numbers (roopa) 12.9% 11.1% 10.5%
code-mixed OpenSLR104 39.4% 36.9% 33.3%
en_clean 3.9% 4.3% 4.5%
en_indian 14.0% 15.4% 15.5%
en_tech 15.4% 14.3% 15.4%
hi_read 12.0% 12.5% 12.1%
hi_unseen 7.1% 7.1% 7.1%
hinglish (mucs) 44.4% 44.4% 42.9%
OVERALL 12.5% 12.5% 12.5%

v12 = best code-mixed performance to date (roopa 10.5%, OpenSLR104 33.3%, hinglish 42.9%) while holding overall at 12.5%.

Usage

Same streaming cache-aware interface as the base Nemotron-3.5-ASR model (prompt conditioning + language-ID, chunk sizes 80/160/320/560/1120 ms, default_prompt_mode unified, unified_auto_ratio 0.9).

Downloads last month
5
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support