canary-mlx 0.1.0

NVIDIA Canary ASR model optimized for Apple Silicon with MLX

Canary MLX

NVIDIA Canary ASR model optimized for Apple Silicon using MLX.

Installation

pip install canary-mlx

Usage

Basic Transcription

from canary_mlx import load_model

# Load the model (downloads from HuggingFace automatically)
model = load_model("your-username/canary-mlx")

# Transcribe an audio file
result = model.transcribe("audio.wav", language="en")
print(result)

Load from Local Directory

from canary_mlx import load_model

# Load from a local model directory
model = load_model("./path/to/model")

With Timestamps

result = model.transcribe("audio.wav", language="en", timestamps=True)

for sentence in result.sentences:
    print(f"[{sentence.start:.2f}s - {sentence.end:.2f}s] {sentence.text}")

Long Audio with Chunking

result = model.transcribe(
    "long_audio.wav",
    language="en",
    timestamps=True,
    chunk_duration=30.0,  # Process in 30-second chunks
    overlap_duration=15.0,  # 15-second overlap between chunks
)

Supported Languages

Bulgarian (bg), Croatian (hr), Czech (cs), Danish (da), Dutch (nl), English (en), Estonian (et), Finnish (fi), French (fr), German (de), Greek (el), Hungarian (hu), Italian (it), Latvian (lv), Lithuanian (lt), Maltese (mt), Polish (pl), Portuguese (pt), Romanian (ro), Slovak (sk), Slovenian (sl), Spanish (es), Swedish (sv), Russian (ru), Ukrainian (uk)

API Reference

load_model(path_or_hf_id, dtype=mx.bfloat16)

Load a Canary model from a local directory or HuggingFace Hub.

Parameters:

• path_or_hf_id: Local path or HuggingFace model ID (e.g., "your-username/canary-mlx")
• dtype: Data type for model weights (default: mx.bfloat16)

Returns: Canary model instance

model.transcribe(...)

Transcribe an audio file.

Parameters:

• path: Path to audio file
• language: Language code (e.g., "en")
• timestamps: Include word-level timestamps (default: False)
• punctuation: Include punctuation (default: True)
• chunk_duration: Process in chunks of this duration (optional)
• overlap_duration: Overlap between chunks in seconds (default: 15.0)

Returns: TranscriptionResult if timestamps=True, else str

Model Conversion

To convert a NeMo Canary model to MLX format:

1. Install conversion dependencies:

   pip install "canary-mlx[convert]"
   

2. Download and extract the NeMo model:

   tar -xvf canary_model.nemo -C canary_untared
   

3. Run the conversion script:

   python convert_nemo.py
   

This creates a model/ directory containing:

• model.safetensors - Model weights
• config.json - Model configuration
• tokenizer.model - SentencePiece tokenizer

License

MIT License