sampathlonka/svarupa_asr_0.6b_v1

Svarupa ASR 0.6B v1 (svarupa_asr_0.6b_v1)

Svarupa ASR 0.6B v1 is a fine-tuned version of Nemotron_3.5_asr_0.6b (nvidia/nemotron-3.5-asr-streaming-0.6b), specialized for Hinglish (Hindi-English code-switching), conversational customer chatbot audio, rural telephony acoustics, and formal Hindi/Indian-English.

The model is built on NVIDIA's FastConformer Transducer (RNNT) cache-aware streaming architecture (600M parameters) and fine-tuned on a 2-node GB10 GPU cluster across multi-stage distributed optimization runs consuming nearly 3,000 cleaned audio hours and over 1.4 million speech cuts.

Model Details

• —Base Pretrained Model: nvidia/nemotron-3.5-asr-streaming-0.6b (NVIDIA)
• —Model Architecture: FastConformer-Transducer (RNNT) with Cache-Aware Streaming
• —Parameter Count: ~600M parameters
• —Sampling Rate: 16 kHz mono PCM
• —Supported Languages & Dialects: Hindi (hi), English (en), Hinglish (hi-en code-switching), 16+ Indian regional dialects
• —Streaming Latency: Cache-aware chunk size = 160ms with left-context caching

Training Data & Cleaned Hours

The model was trained on a curated multi-slice acoustic mix totaling 2,945.4 cleaned audio hours and 1,436,901 audio cuts across diverse real-world domains:

Training History & Step Strategy

• —Initial Optimization Phase: ~71,000 – 85,000 sample-based steps on the primary 815,105-utterance corpus to establish initial transducer alignment.
• —Duration-Batched Multi-Node Phase (V2 Pipeline): 20,000 cumulative steps across 2 nodes (8 H100 GPUs) using Lhotse dynamic duration batching (batch_duration=240s per GPU = 480 seconds / 8 minutes of speech per step):
• —Run 5: 5,000 steps establishing code-switching foundation.
• —Run 6: 5,000 steps incorporating Whisper-cleaned chatbot audio.
• —Run 7 (Final): 10,000 steps (9 full epochs) ingesting rural telephony acoustics and synthetic pairs.
• —Total Audio Throughput: Over 12,000 cumulative audio hours processed across optimization passes.

Key Performance Benchmarks

All evaluations performed with Lever 2 (v7 Normalizer) and Lever 1 (Beam Search Decoding, Beam=4):

Quickstart & Inference

1. Installation

pip install nemo_toolkit['asr'] torch torchaudio

2. Loading the Model with NeMo

import nemo.collections.asr as nemo_asr
from omegaconf import open_dict

# Restore the model
model = nemo_asr.models.ASRModel.restore_from("svarupa_asr_0.6b_v1.nemo")
model.eval()

# Configure Beam Search Decoding (Beam Size = 4)
cfg = model.cfg.decoding
with open_dict(cfg):
    cfg.strategy = "malsd_batch"
    if "beam" not in cfg:
        cfg.beam = {}
    cfg.beam.beam_size = 4
    cfg.beam.max_symbols = 35
    cfg.beam.return_best_hypothesis = True
    cfg.beam.score_norm = True
    cfg.beam.allow_cuda_graphs = False

model.change_decoding_strategy(cfg)

# Transcribe audio file (16kHz mono WAV recommended)
transcriptions = model.transcribe(["sample.wav"], batch_size=4)
print("Transcription:", transcriptions[0])

Intended Use & Limitations

• —Intended Use: Real-time conversational AI, voice bots, telephony voice assistants, customer service automation, and transcription of bilingual Indian speech.
• —Input Requirements: 16,000 Hz single-channel audio.
• —Code-Switching Capabilities: Handles seamless switching between English and Hindi within the same sentence (matrix language Hindi with embedded English nouns, verbs, and tech terminology).

License & Attribution

• —Base Model: nvidia/nemotron-3.5-asr-streaming-0.6b (NVIDIA)
• —Fine-Tuned by: Svarupa Team
• —License: Apache 2.0
• —Citation: Please cite Svarupa ASR 0.6B v1 when using this model in academic or commercial applications.