indicconformer-120m-onnx

IndicConformer 120M — Per-Language ONNX Models for Indian Language ASR

CTC-only ONNX exports of AI4Bharat's IndicConformer hybrid CTC/RNN-T large models (~120M parameters each). One model per language for 12 Indian languages, optimized for batch inference with ONNX Runtime (GPU via CUDA/ROCm, or CPU).

Languages

Repo Structure

{lang}/model.onnx    # ONNX model (~470 MB each)
{lang}/vocab.json    # BPE vocabulary (JSON list of token strings)

Model Details

Mel Spectrogram Preprocessing

These models require NeMo-compatible preprocessing. The full pipeline:

1. Preemphasis — coefficient 0.97
2. STFT — n_fft=512, hop_length=160, win_length=400
3. Mel filterbank — 80 bins, Slaney normalization, fmin=0, fmax=8000
4. Power — power=2.0 (power spectrogram)
5. Log — log(x + 2^-24)
6. Per-feature normalization — zero mean, unit variance per mel bin (unbiased std, ddof=1)

import numpy as np
import librosa

def nemo_mel_spectrogram(audio: np.ndarray, sr: int = 16000) -> np.ndarray:
    """NeMo-compatible log-mel spectrogram. Returns float32 [80, T_frames]."""
    # 1. Preemphasis
    audio = np.concatenate([audio[:1], audio[1:] - 0.97 * audio[:-1]])

    # 2. Mel power spectrogram
    mel = librosa.feature.melspectrogram(
        y=audio, sr=sr, n_fft=512, hop_length=160, win_length=400,
        n_mels=80, fmin=0, fmax=sr / 2, norm="slaney", power=2.0,
    )

    # 3. Log
    log_mel = np.log(mel + 2**-24).astype(np.float32)

    # 4. Per-feature normalization
    mean = log_mel.mean(axis=1, keepdims=True)
    std  = log_mel.std(axis=1, ddof=1, keepdims=True) + 1e-5
    log_mel = (log_mel - mean) / std

    return log_mel.astype(np.float32)

Quick Start

import json, numpy as np, librosa, onnxruntime as ort
from huggingface_hub import hf_hub_download

REPO = "sulabhkatiyar/indicconformer-120m-onnx"
LANG = "hi"  # Change to any supported language code

# 1. Download model and vocabulary
model_path = hf_hub_download(REPO, f"{LANG}/model.onnx")
vocab_path = hf_hub_download(REPO, f"{LANG}/vocab.json")

# 2. Load ONNX session (auto-selects best available provider)
providers = ["CUDAExecutionProvider", "ROCMExecutionProvider", "CPUExecutionProvider"]
session = ort.InferenceSession(model_path, providers=providers)

# 3. Load vocabulary
with open(vocab_path, "r", encoding="utf-8") as f:
    vocab = json.load(f)
blank_id = len(vocab)

# 4. Load and preprocess audio
audio, sr = librosa.load("audio.wav", sr=16000)

# Preemphasis
audio_pe = np.concatenate([audio[:1], audio[1:] - 0.97 * audio[:-1]])

# Mel spectrogram
mel = librosa.feature.melspectrogram(
    y=audio_pe, sr=16000, n_fft=512, hop_length=160, win_length=400,
    n_mels=80, fmin=0, fmax=8000, norm="slaney", power=2.0,
)
log_mel = np.log(mel + 2**-24).astype(np.float32)
mean = log_mel.mean(axis=1, keepdims=True)
std  = log_mel.std(axis=1, ddof=1, keepdims=True) + 1e-5
log_mel = (log_mel - mean) / std

# Batch dimensions: [1, 80, T]
mel_batch  = log_mel[np.newaxis, :, :].astype(np.float32)
mel_length = np.array([mel_batch.shape[2]], dtype=np.int64)

# 5. Run inference
logits = session.run(None, {"audio_signal": mel_batch, "length": mel_length})[0]

# 6. CTC greedy decode
log_probs = logits - logits.max(axis=-1, keepdims=True)
log_probs = log_probs - np.log(np.exp(log_probs).sum(axis=-1, keepdims=True))
ids = np.argmax(log_probs[0], axis=-1)

# Collapse consecutive duplicates, remove blanks
prev, tokens = -1, []
for t in ids:
    if t != prev:
        if t != blank_id and t < len(vocab):
            tokens.append(vocab[t])
    prev = t

transcript = "".join(tokens).replace("\u2581", " ").strip()
print(transcript)

Batch Inference

For throughput, pad mel spectrograms to the same length within a batch and set the length input to each utterance's actual frame count. The model handles variable-length inputs natively — invalid frames beyond the given length are ignored.

# Assuming mels is a list of [80, T_i] arrays
mel_lengths = np.array([m.shape[1] for m in mels], dtype=np.int64)
max_len = int(mel_lengths.max())
mel_batch = np.zeros((len(mels), 80, max_len), dtype=np.float32)
for i, m in enumerate(mels):
    mel_batch[i, :, :m.shape[1]] = m

logits = session.run(None, {"audio_signal": mel_batch, "length": mel_lengths})[0]
# Decode each utterance using its actual output length (~T_mel / 4)

Export Process

These ONNX models were exported from AI4Bharat's NeMo .nemo checkpoints:

1. Load the hybrid CTC/RNN-T model: EncDecHybridRNNTCTCBPEModel.restore_from(nemo_path)
2. Switch to CTC decoder: model.cur_decoder = "ctc"
3. Export: model.export("model.onnx")
4. Extract vocabulary: list(model.ctc_decoder.vocabulary) saved as vocab.json

Exported using AI4Bharat's NeMo fork (branch nemo-v2).

Credits

• AI4Bharat for training and releasing the original IndicConformer models
• Original model repos: ai4bharat/indicconformer_stt_{lang}_hybrid_ctc_rnnt_large on HuggingFace