senstella/parakeet-mlx

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

What it solves

It provides a way to run Nvidia's Parakeet Automatic Speech Recognition (ASR) models on Apple Silicon hardware using the MLX framework, enabling efficient local audio transcription.

How it works

The project implements the Parakeet model architecture in MLX, allowing users to load pre-trained weights from Hugging Face. It supports various model variants (TDT, RNNT, CTC, TDTCTC) and offers multiple decoding methods, such as greedy and beam search. To handle long audio files, it includes chunking mechanisms and local attention to reduce memory usage.

Who it’s for

Developers and users with Apple Silicon Macs who need high-performance, local speech-to-text transcription with support for timestamps and streaming audio.

Highlights

  • Multiple Output Formats: Supports exporting transcriptions as TXT, SRT, VTT, and JSON.
  • Precise Timestamps: Provides word-level and sentence-level alignments for audio.
  • Streaming Support: Includes a transcribe_stream method for real-time transcription.
  • Flexible Decoding: Offers configurable beam search parameters (beam size, length penalty, patience) for TDT models.
  • Memory Optimization: Implements local attention to handle long audio sequences without excessive memory consumption.

Related

  • Project
  • Project
  • Project
  • Project