Quds-v4-onnx

🚨 Non Commercial Usage Only

💡 Fine-tuned specifically for the domain of Islamic lectures and specialized Howzah courses, such as Tafsir, Fiqh, Usul, and Rijal.

🔔 FOR NeMo version (includes both rnnt and ctc decoders and word timestamps) send a request in Community

Model Details

• Model Name: Quds-v4-onnx
• Model Type: Automatic Speech Recognition (ASR) / Speech-to-Text
• Architecture: NeMo FastConformer Large Hybrid (RNN-T variant)
• Format: ONNX (Open Neural Network Exchange)
• Language: Persian (Farsi)

Model Description

Quds-v4-onnx is a highly accurate Persian Automatic Speech Recognition model. It is based on the robust NeMo FastConformer Large Hybrid architecture and is specifically exported in the ONNX format to enable fast, cross-platform inference. This release features an RNN-T (Recurrent Neural Network Transducer) decoder.

Performance & Accuracy

• The model achieves very high accuracy on standard, formal Persian speech.

Training Details

• Training Dataset: The model was trained on a high-quality, non-public (private) dataset.
• Dataset Size: 600 hours of audio data.

---

Usage

This model can be used seamlessly with the onnx-asr Python library, which provides a clean interface for ONNX-based ASR models.

Installation

First, install the onnx-asr package:

pip install onnx-asr

1. Basic Inference (Single File)

Load the model directly from Hugging Face and transcribe a .wav file:

import onnx_asr

# Load the model from Hugging Face
model = onnx_asr.load_model("hojreh/Quds-v4-onnx")

# Recognize speech from a single WAV file
transcription = model.recognize("test.wav")
print(transcription)

2. Batch Processing

You can pass a list of file paths to process multiple audio files efficiently in a single batch:

import onnx_asr

model = onnx_asr.load_model("hojreh/Quds-v4-onnx")
transcriptions = model.recognize(["test1.wav", "test2.wav", "test3.wav"])
print(transcriptions)

3. Long audios (VAD Integration)

For long audio files, you can integrate a VAD model (such as Silero) to skip non-speech segments:

import onnx_asr

# Load a VAD model
vad = onnx_asr.load_vad("onnx-community/pyannote-segmentation-3.0") # or "silero"

# Attach VAD to the ASR model
model = onnx_asr.load_model("hojreh/Quds-v4-onnx").with_vad(vad)

# Recognize with VAD (returns segmented results)
for res in model.recognize("test.wav"):
    print(res)

4. Timestamps

import onnx_asr
model = onnx_asr.load_model("hojreh/Quds-v4-onnx").with_timestamps()
print(model.recognize("test1.wav"))

5. Quantized version (INT8)

Quantized version (INT8) is smaller but NOT accurate as main version.

import onnx_asr
model = onnx_asr.load_model("hojreh/Quds-v4-onnx", quantization="int8")
print(model.recognize("test1.wav"))

6. Command Line Interface (CLI)

The onnx-asr package also includes a simple CLI for quick testing directly from your terminal:

# Run transcription on a wav file
onnx-asr hojreh/Quds-v4-onnx test.wav

(For full CLI parameters and options, run onnx-asr -h in your terminal)

---

Limitations

1. Dialect Sensitivity: Due to training on standard Persian audio, the model struggles with heavy regional accents, colloquialisms, and non-standard dialects of Persian.
2. Audio Quality Sensitivity: As with most ASR models, performance degrades in highly noisy environments.
3. Domain Specificity: The model may not be perfectly suited for highly specialized technical or medical domains.