nemotron-hinglish-v2

Nemotron Hinglish v2

Nemotron-Hinglish-v2 is a continued 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.

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

Continue fine-tuned on a much larger bilingual + code-mixed mix (~4,726 h across 2.05M utterances):

• Trained with punctuation/casing preserved (matching base-model text style) using NeMo's prompt-conditioned RNNT (EncDecRNNTBPEModelWithPrompt).
• Training prompt mode mixes forced langID with auto (language self-detection, 0.9 ratio) to handle code-switching.

Benchmark

Measured on held-out clips, auto prompt, greedy decode, punctuation-insensitive WER (500 utts/language sample):

Usage (NeMo)

import nemo.collections.asr as nemo_asr
model = nemo_asr.models.ASRModel.restore_from("nvidia/nemotron-hinglish-v2")
model.eval()
transcriptions = model.transcribe(["audio.wav"], batch_size=4)
print(transcriptions)