WeST

Implement speech transcription based on LLM in just 300 lines of code.

CommonProductProgrammingSpeech RecognitionNatural Language Processing

WeST is an open-source speech recognition transcription model that achieves speech-to-text conversion in a concise format of 300 lines of code, based on a large language model (LLM). It includes a large language model, a speech encoder, and a projector, with only the projector being trainable. The development of WeST is inspired by SLAM-ASR and LLaMA 3.1, aiming to deliver efficient speech recognition capabilities through simplified code.

Integrates interchangeable large language models like LLaMA or QWen.
Uses speech encoders
such as Whisper
to encode speech signals.
Supports jsonl format configuration for custom training and testing data.
Provides detailed configuration options for training parameters
including learning rate
weight decay
etc.
Supports Deepspeed configuration to optimize the model training process.

WeST primarily targets developers and data scientists
especially professionals interested in the fields of speech recognition and natural language processing. Its simplicity and ease of use make it an ideal choice for rapid prototyping and academic research.

Developers quickly build prototypes of voice assistants using WeST.
Researchers conduct experiments and write papers on speech recognition technology with WeST.
Educational institutions use WeST as a teaching tool to demonstrate how speech recognition works.

1. Prepare training and testing datasets
ensuring they meet the jsonl format requirements.
2. Set up the Python environment and install necessary dependencies according to project requirements.
3. Configure training parameters
including learning rate

Visit

WeST Visit Over Time

Monthly Visits

494758773

Bounce Rate

37.69%

Page per Visit

5.7

Visit Duration

00:06:29

WeST Visit Trend

WeST Visit Geography

WeST Traffic Sources

WeST Alternatives

Q-RWKV-6 32B Instruct Preview — The most powerful variant of the RWKV model, breaking multiple English benchmark tests.

Programming•Machine Learning•Natural Language Processing

OLMo-2-1124-7B-SFT — High-performance English text generation model

Productivity•Text Generation•Natural Language Processing

Llama-3.1-Tulu-3-8B-DPO — An advanced text generation model that supports diverse tasks.

Programming•Text Generation•Natural Language Processing

RWKV v6 Finch 14B — RWKV v6 Finch 14B is an open-source large model that efficiently processes long texts.

Productivity•Artificial Intelligence•Natural Language Processing

234

OpenVoiceChat — Engage in natural voice conversations with large language models.

chatting•Speech Recognition•Text-to-Speech

282

Llama3-s v0.2 — Latest multimodal checkpoint to enhance speech comprehension capabilities.

Programming•Speech Recognition•Natural Language Processing

186

1.5-Pints — A compact large language model pre-trained in 9 days

Programming•Open Source•Machine Learning

168

speech-to-speech — Open-source speech-to-speech conversion module

Programming•Speech Recognition•Natural Language Processing

540

llama3-s — An open-source language model currently being trained, equipped with 'hearing' capabilities.

Programming•Natural Language Processing•Machine Learning

102

Agents 2.0 — Open-source framework supporting data-driven adaptive language agents.

Programming•Open Source•Natural Language Processing

Gemini 1.5 Flash — A lightweight and high-performance AI model from Google, designed for large-scale, high-frequency tasks.

Productivity•Machine Learning•Multimodal

522

MAP-NEO — MAP-NEO is an entirely open-source large language model offering advanced natural language processing capabilities.

Programming•Natural Language Processing•Open Source

438

Granite Code Models — An open-source foundation model designed for code intelligence tasks, supporting 116 programming languages.

Programming•Code Intelligence•Machine Learning

300

Cognita — An open-source framework for constructing modular production-level applications.

Programming•Machine Learning•Natural Language Processing

192

Qwen1.5-110B — The first petabyte-scale open-source model in the Qwen1.5 series, supporting multilingual functionality and an efficient Transformer decoder architecture.

ChineseSelection•Artificial Intelligence•Machine Learning

570

Snowflake Arctic — An efficient enterprise-level AI model that achieves high-quality custom models at a low cost.

InternationalSelection•Artificial Intelligence•Machine Learning

168

OpenELM — OpenELM is a family of efficient language models equipped with an open-source training and inference framework.

InternationalSelection•Language Model•Natural Language Processing

714

Docker GenAI Stack — An AI application development solution launched by Docker

Programming•Docker•Machine Learning

234

OmniAudio-2.6B — The fastest edge-deployed audio language model in the world.

Productivity•Audio Processing•Edge Computing

CosyVoice Speech Generation Model 2.0-0.5B — Efficient, multilingual speech synthesis model

Music•Speech Synthesis•Artificial Intelligence

114

tulu-3-sft-olmo-2-mixture — A large-scale multilingual text dataset.

Others•Multilingual•Text Dataset

OLMo-2-1124-7B-RM — A large language model for text generation and classification.