Saboorhsn/quran-stt-onnx

Zipformer-v3 Arabic / Quranic ASR & Forced Alignment

Attribution & Credits:
This model family is based on the state-of-the-art Zipformer2-CTC architecture developed by Quran-Lab (Quran-Lab/zipformer_p-arabic-v3). All original training, phonetization, and dataset curation credits belong to the Quran-Lab team and the authors of quran-transcript.

---

Quantization & Model Variants Table

This repository hosts ONNX quantized and precision-optimized variants designed for edge, CPU, and high-throughput GPU forced-alignment pipelines:

Model File	Precision / Quant	Size	Recommended Platform	Realtime Speedup (RTF)
zipformer_p_arabic_v3.1.fp16.onnx	Pure FP16 (Tensor Core)	126.2 MB	NVIDIA GPU (CUDA / Tensor Cores)	>220x – 235x realtime (RTF < 0.0045)
zipformer_p_arabic_v3.1.onnx	FP32 (v3.1 madd fine-tune)	250.8 MB	High-precision CPU / GPU	~100x realtime (GPU) / 3.5x (CPU)
zipformer_p_arabic_v3.1.int8.onnx	INT8 / Q8 Quantized	69.3 MB	Mobile / Embedded / Laptop CPU	~12x – 15x realtime (x86 CPU)
zipformer_p_arabic_v3.onnx	FP32 (v3 Base)	250.8 MB	Reference FP32	~100x realtime (GPU)
zipformer_p_arabic_v3.int8.onnx	INT8 / Q8 (v3 Base)	69.3 MB	Standard CPU deployment	~12x realtime (CPU)
tokens.txt	CTC Symbol Table (251 units)	2.3 KB	Required for decoding	—
config.json	Model Hyperparameters	177 B	Config metadata	—

---

Benchmark Highlights (Surah Al-Kahf 18 & Surah Al-Baqarah 2)

Evaluated across continuous full-surah recitations on NVIDIA Tesla T4 GPU:

• Parity: 0.00 ms mean delta against FP32 reference alignment.
• Throughput: Aligns full 24.8-minute continuous audio (1,584 words) in 6.69 seconds flat.
• Entire 30-Juz Quran (~25 hours audio): Full word-level forced-alignment and repeat-detection in ~6.3 minutes.

---

Usage with quran-forced-align

Install and align with zero configuration:

# Installs and runs with automatic GPU/CPU detection and auto-model caching:
pip install quran-forced-align

# Single Surah alignment (JSON + SRT):
quran-forced-align --surah 66 --audio audio/066_basit.mp3 --out output/066.json

Python Streaming ONNX Inference

import onnxruntime as ort
import numpy as np

# Load FP16 model on CUDA Execution Provider:
opts = ort.SessionOptions()
opts.graph_optimization_level = ort.GraphOptimizationLevel.ORT_ENABLE_ALL

session = ort.InferenceSession(
    "zipformer_p_arabic_v3.1.fp16.onnx",
    sess_options=opts,
    providers=[('CUDAExecutionProvider', {'device_id': 0, 'enable_cuda_graph': True}), 'CPUExecutionProvider']
)

---

Original Model Tree Files

Included in this repository are all symbol mappings and reference metadata from the original release:

• tokens.txt: CTC 251-symbol table.
• config.json: Model configuration.

License

Released under the Quran-Lab No-Profit License, Version 1.1 (NPL-1.1); free to use, study, adapt and redistribute for non-profit and educational purposes.