eelcor/canary-1b-v2-mlx

Canary 1B v2 MLX

This is a conversion of NVIDIA's Canary 1B v2 ASR model to Apple's MLX framework for fast inference on Apple Silicon.

Model Description

Canary 1B v2 is a state-of-the-art automatic speech recognition (ASR) model from NVIDIA, featuring:

• Architecture: FastConformer Encoder (32 layers) + Transformer Decoder (8 layers)
• Parameters: ~1B
• Languages: 25 languages including Dutch, English, German, French, Spanish, and more
• Training Data: NVIDIA Granary dataset

Performance

Benchmarked on Apple Silicon (M-series):

Metric	Value
Speed	~8.5x real-time
RTF (Real-Time Factor)	~0.12

Comparison with Whisper (10s Dutch audio):

Model	Speed	Dutch Quality
Canary 1B v2 MLX	8.5x RT	Excellent
Whisper large-v2 MLX	6.4x RT	Excellent

Canary is 33% faster than Whisper with comparable quality.

Usage

Installation

pip install canary-mlx

Python API

from canary_mlx import load_model

# Load the model
model = load_model("eelcor/canary-1b-v2-mlx")

# Transcribe audio (specify language for best results)
result = model.transcribe("audio.wav", language="nl")  # Dutch
print(result)

# For English
result = model.transcribe("audio.wav", language="en")

Supported Languages

The model supports 25 languages. Use the ISO 639-1 language codes:

• nl - Dutch
• en - English
• de - German
• fr - French
• es - Spanish
• it - Italian
• pt - Portuguese
• pl - Polish
• ru - Russian
• uk - Ukrainian
• And more...

Conversion Details

This model was converted from the original NeMo format using:

• PyTorch checkpoint extraction
• Weight tensor transposition for MLX conv layers
• SentencePiece tokenizer preservation

License

This model inherits the CC-BY-4.0 license from the original NVIDIA Canary model.

Citation

If you use this model, please cite the original NVIDIA work:

@misc{nvidia_canary_2024,
  title={Canary: A Multilingual ASR Model},
  author={NVIDIA NeMo Team},
  year={2024},
  publisher={NVIDIA}
}

Acknowledgments

• NVIDIA NeMo Team for the original Canary model
• canary-mlx library for MLX inference support
• Apple MLX framework