Nemotron Hinglish v3

Nemotron-Hinglish-v3 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 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 for code-mixing)
  • Streaming: cache-aware, chunk sizes 80/160/320/560/1120 ms

Training Data

Fine-tuned on a large bilingual + code-mixed mix with number/symbol-dense corpora added for digit & phone-number accuracy:

Slice Language Role
SPGISpeech en clean English
IISc_SPICOR en Indian-accent English
Peoples Speech en large English
SPRING Hindi-1482Hrs + Shrutilipi hi Hindi
IndicVoices-R Hindi hi Hindi
UJS + Hinglish-CC hinglish code-mixed
FLEURS (en + hi) en/hi ~21% digit-dense
earnings22 en ~22% digit, ~7.7% symbol — phone/number rich

Trained with auto language self-detection (0.9 ratio) for code-switching, punctuation/casing preserved.

Benchmark

Greedy decode, punctuation-insensitive WER (held-out sample, auto prompt). Compared to previous releases:

Language Nemotron-Hinglish-v3 v2 v1
English 4.0% 5.1% 4.2%
Hindi 12.2% 13.4% 12.4%
Hinglish 29.2% 30.0% 22.6%
  • Digit / phone-number accuracy improved ~32% (fewer digit-sequence errors than v2), from added number-dense corpora.

Usage (NeMo)

import nemo.collections.asr as nemo_asr
model = nemo_asr.models.ASRModel.restore_from("nvidia/nemotron-hinglish-v3")
model.eval()
transcriptions = model.transcribe(["audio.wav"], batch_size=4)
print(transcriptions)
Downloads last month
17
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for smajji/nemotron-hinglish-v3

Finetuned
(51)
this model