Skip to content

WeST

Speech-to-text transcription in about 300 lines of code

Visit official site

What WeST is

The model is described as an open-source speech recognition system that turns speech into text with a large language model in about 300 lines of code. It consists of an LLM, a speech encoder and a projector.

What you can do with it

  • Swap the LLM for one such as LLaMA or QWen
  • Encode speech with an encoder such as Whisper
  • Configure training and test data in JSONL format

Application scenarios

  • A developer builds a speech assistant prototype
  • A researcher runs recognition experiments
  • A teaching team demonstrates the mechanism

Pros & cons

✓ What we like

  • Training parameters include learning rate and weight decay
  • Deepspeed configuration is supported

! What to watch out for

  • Only the projector part is trainable
  • Training and test data must follow the JSONL format

FAQ

What is trainable?

The material states only the projector part can be trained.

What inspired the project?

The material names SLAM-ASR and LLaMA 3.1 as the inspiration.

Last reviewed: 2026-11-15

More AI audio tools tools

View all →

How we review