parakeet-mlx 0.2.6

An implementation of the Nvidia's Parakeet models for Apple Silicon using MLX.

Parakeet MLX

An implementation of the Parakeet models - Nvidia's ASR(Automatic Speech Recognition) models - for Apple Silicon using MLX.

Installation

[!NOTE] Make sure you have ffmpeg installed on your system first, otherwise CLI won't work properly.

Using uv - recommended way:

uv add parakeet-mlx -U

Or, for the CLI:

uv tool install parakeet-mlx -U

Using pip:

pip install parakeet-mlx -U

CLI Quick Start

parakeet-mlx <audio_files> [OPTIONS]

Arguments

• audio_files: One or more audio files to transcribe (WAV, MP3, etc.)

Options

• --model (default: mlx-community/parakeet-tdt-0.6b-v2)

  • Hugging Face repository of the model to use

• --output-dir (default: current directory)

  • Directory to save transcription outputs

• --output-format (default: srt)

  • Output format (txt/srt/vtt/json/all)

• --output-template (default: {filename})

  • Template for output filenames, {filename}, {index}, {date} is supported.

• --highlight-words (default: False)

  • Enable word-level timestamps in SRT/VTT outputs

• --verbose / -v (default: False)

  • Print detailed progress information

• --chunk-duration (default: 120 seconds)

  • Chunking duration in seconds for long audio, 0 to disable chunking

• --overlap-duration (default: 15 seconds)

  • Overlap duration in seconds if using chunking

• --fp32 / --bf16 (default: bf16)

  • Determinate the precision to use

Examples

# Basic transcription
parakeet-mlx audio.mp3

# Multiple files with word-level timestamps of VTT subtitle
parakeet-mlx *.mp3 --output-format vtt --highlight-words

# Generate all output formats
parakeet-mlx audio.mp3 --output-format all

Python API Quick Start

Transcribe a file:

from parakeet_mlx import from_pretrained

model = from_pretrained("mlx-community/parakeet-tdt-0.6b-v2")

result = model.transcribe("audio_file.wav")

print(result.text)

Check timestamps:

from parakeet_mlx import from_pretrained

model = from_pretrained("mlx-community/parakeet-tdt-0.6b-v2")

result = model.transcribe("audio_file.wav")

print(result.sentences)
# [AlignedSentence(text="Hello World.", start=1.01, end=2.04, duration=1.03, tokens=[...])]

Do chunking:

from parakeet_mlx import from_pretrained

model = from_pretrained("mlx-community/parakeet-tdt-0.6b-v2")

result = model.transcribe("audio_file.wav", chunk_duration=60 * 2.0, overlap_duration=15.0)

print(result.sentences)

Timestamp Result

• AlignedResult: Top-level result containing the full text and sentences
  • text: Full transcribed text
  • sentences: List of AlignedSentence
• AlignedSentence: Sentence-level alignments with start/end times
  • text: Sentence text
  • start: Start time in seconds
  • end: End time in seconds
  • duration: Between start and end.
  • tokens: List of AlignedToken
• AlignedToken: Word/token-level alignments with precise timestamps
  • text: Token text
  • start: Start time in seconds
  • end: End time in seconds
  • duration: Between start and end.

Low-Level API

To transcribe log-mel spectrum directly, you can do the following:

import mlx.core as mx
from parakeet_mlx.audio import get_logmel, load_audio

# Load and preprocess audio manually
audio = load_audio("audio.wav", model.preprocessor_config.sample_rate)
mel = get_logmel(audio, model.preprocessor_config)

# Generate transcription with alignments
# Accepts both [batch, sequence, feat] and [sequence, feat]
# `alignments` is list of AlignedResult. (no matter you fed batch dimension or not!)
alignments = model.generate(mel)

Todo

• Add CLI for better usability
• Add support for other Parakeet varients
• Streaming input (Although RTFx is MUCH higher than 1 currently - it should be much sufficient to stream with current state)
• Option to enhance choosen words' accuracy
• Chunking with continuous context (I think it might be able to achieve by preserving decoder state - just a speculation though)

Acknowledgments

• Thanks to Nvidia for training this awesome models and writing cool papers and providing nice implementation.
• Thanks to MLX project for providing the framework that made this implementation possible.
• Thanks to audiofile and audresample, numpy, librosa for audio processing.
• Thanks to dacite for config management.

License

Apache 2.0