Note: Models successfully exported and validated with ONNX Runtime. Python verification with sherpa-onnx pending due to cache dimension configuration. On-device testing recommended.

Configuration

Streaming Parameters

Attention Context: [128, 0] (left, right)
Attention Model: rel_pos_local_attn
Convolution Context: causal
Window Size: 128 frames
Chunk Shift: 1

Model Architecture

Vocabulary: 1024 tokens (BPE)
Sample Rate: 16000 Hz
Feature Dimension: 80 (log-Mel)
Subsampling Factor: 8x
Prediction Layers: 1
Prediction Hidden: 640

Files

File	Size	Description
`encoder.int8.onnx`	~119 MB	INT8 quantized streaming encoder
`decoder.int8.onnx`	~5 MB	INT8 quantized prediction network
`joiner.int8.onnx`	~5 MB	INT8 quantized joiner network
`tokens.txt`	~13 KB	1024-token BPE vocabulary
`meta.json`	<1 KB	Model configuration and metadata
`verification_report.json`	<1 KB	Detailed verification results

Total Download: ~129 MB

Usage

Python (sherpa-onnx)

python

import sherpa_onnx

# Create recognizer
recognizer = sherpa_onnx.OnlineRecognizer.from_transducer(
    encoder="encoder.int8.onnx",
    decoder="decoder.int8.onnx",
    joiner="joiner.int8.onnx",
    tokens="tokens.txt",
    sample_rate=16000,
    feature_dim=80,
    decoding_method="greedy_search"
)

# Transcribe audio (16kHz, mono, float32)
stream = recognizer.create_stream()
stream.accept_waveform(16000, audio_samples)
stream.input_finished()

while recognizer.is_ready(stream):
    recognizer.decode_stream(stream)

result = recognizer.get_result(stream)
print(result)  # Quranic Arabic transcription with diacritics

Mobile (Android/iOS)

See sherpa-onnx mobile documentation for integration guide.

Key Integration Points:

Download all 4 files (encoder, decoder, joiner, tokens)
Initialize OnlineRecognizer with file paths
Feed 16kHz mono audio to recognizer
Get real-time transcriptions with diacritics

Training Details

Fine-Tuning Approach

Method: Progressive layer unfreezing (3 phases)
Dataset: tarteel-ai/everyayah (Quranic recitations)
Training Samples: 167,908
Validation Samples: 20,976
Final NeMo WER: 0.14% (Phase 3)

Streaming Configuration

Applied streaming attention from Phase 1
Causal convolution padding
Left context: 128 frames (~1.28s lookback)
Right context: 0 frames (fully causal)

Export and Quantization

Exported with NeMo cache-aware streaming
Dynamic INT8 quantization (per-channel weights)
~71% size reduction (456MB → 132MB encoder)
Export validated with ONNX Runtime

Verification

⚠️ Pending On-Device Testing

Models have been:

✅ Successfully exported from Phase 3 NeMo checkpoint
✅ Quantized to INT8 with ONNX Runtime
✅ Validated to load correctly in ONNX Runtime
✅ Confirmed to have proper metadata structure

Next step: Test with sherpa-onnx on Android/iOS device to measure actual WER.

See EXPORT_SUCCESS_SUMMARY.md for details on export process.

Limitations

Domain-specific: Optimized for Quranic Arabic recitations
Diacritic-sensitive: Vocabulary includes Arabic diacritics
Streaming only: Exported for real-time inference (not offline batch)
Greedy decoding: For beam search, use NeMo model directly

Citation

bibtex

@misc{fastconformer-quran-ar-onnx-int8,
  author = {Mohammed},
  title = {FastConformer Quranic Arabic ONNX INT8},
  year = {2026},
  publisher = {Hugging Face},
  url = {https://huggingface.co/mohammed/fastconformer-quran-ar-onnx-int8}
}

License

CC-BY-4.0

Acknowledgments

Base model: NVIDIA NeMo FastConformer
Dataset: tarteel-ai/everyayah
Runtime: sherpa-onnx by k2-fsa
Training framework: NVIDIA NeMo

Model Card Contact: Mohammed (via HuggingFace)