mohammedaly22/Metro-ASR-Small

Metro-ASR Small (61M)

Non-Autoregressive CTC Speech Recognition for Egyptian Arabic + Code-Switching

Model Description

Metro-ASR Small is a 61M parameter non-autoregressive CTC-based ASR model built on a modern Conformer encoder. It is specifically designed for Egyptian Arabic (العامية المصرية) with native Arabic-English code-switching support.

Architecture

Component	Details
Encoder	Conformer (12 layers, d_model=384, 6 heads)
Position Encoding	RoPE (Rotary Position Embeddings)
Feed-Forward	SwiGLU (Macaron-style dual FFN)
Normalization	RMSNorm (Pre-norm)
Convolution	SE-Gated Depthwise Separable (kernel=31)
Regularization	Stochastic Depth (rate=0.05)
Auxiliary Loss	Intermediate CTC at layer 6
Tokenizer	BPE (SentencePiece, vocab=5000)
Decoding	CTC Greedy / Beam Search + KenLM
Parameters	61.6M

Performance

Split	WER (%)	CER (%)
All	46.85	28.41
Arabic Only	37.24	17.45
Code-Switching	36.32	17.44

Speed: RTF ~0.002 on CPU (500x faster than real-time)

Usage

Install

pip install metro-asr

Quick Start

from metro_asr import MetroASREngine

engine = MetroASREngine.from_pretrained("small")  # Auto-downloads this model
result = engine.transcribe("audio.wav")
print(result.text)

With Language Model (Beam Search)

engine = MetroASREngine.from_pretrained(
    "small",
    lm_path="lm_5gram.bin",     # Download from this repo
    beam_width=100,
    lm_alpha=0.5,
    lm_beta=5.0,
)

result = engine.transcribe("audio.wav", beam_search=True)
print(result.text)

Batch Transcription

results = engine.transcribe_batch(["audio1.wav", "audio2.wav", "audio3.wav"])
for r in results:
    print(f"{r.text}  (RTF={r.rtf:.4f})")

Files in this Repository

File	Description	Size
model.pt	Model checkpoint (weights + optimizer state)	~706MB
config.yaml	Model architecture configuration	<1KB
bpe.model	SentencePiece BPE tokenizer model	~200KB
bpe.vocab	BPE vocabulary file	~100KB
lm_5gram.bin	KenLM 5-gram language model (optional)	~6GB

Training

Trained on a combination of open-source Egyptian Arabic datasets and custom YouTube data:

• Audio data: 130K+ samples from 8 datasets + YouTube Egyptian Arabic content
• Text data: 1.9M Egyptian sentences for LM training
• Training: 440K steps, batch size 32, 4x gradient accumulation, AdamW + Cosine LR
• Hardware: Single GPU training

Limitations

• Optimized for Egyptian Arabic dialect; MSA and other dialects may have higher error rates
• Code-switching support is Arabic-English only
• Best performance on audio 0.5s-30s in duration
• The language model significantly improves accuracy but adds ~6GB to download

Citation

@software{metro-asr-2025,
  title  = {Metro-ASR: Non-Autoregressive CTC-based ASR for Egyptian Arabic and Code-Switching},
  author = {Mohammed Aly},
  year   = {2025},
  url    = {https://github.com/mohammedaly22/Metro-ASR}
}

License

MIT