Audio Models

Liquid’s LFM audio models are among the smallest fully interleaved audio/text-in, audio/text-out models with a complete reasoning backbone — eliminating the need to combine separate TTS/ASR encoders with a standalone language model.

Text-to-Speech

Natural speech synthesis with multiple voice styles and personalities.

Speech Recognition

Multilingual automatic speech recognition and transcription.

Audio-to-Audio

Fully interleaved voice chat with reasoning and generation in one model.

Audio Function Calling

Voice-driven tool use, agent workflows, and API orchestration.

LFM2.5 Models   Latest release

LFM2.5-Audio adds a custom LFM-based audio detokenizer, llama.cpp-compatible GGUFs for CPU inference, and improved ASR and TTS performance over LFM2-Audio.

LFM2.5-Audio-1.5B

1.5B · Recommended

Best audio model for most use cases. Fast, accurate, and CPU-friendly.

LFM2 Models

LFM2-Audio-1.5B

1.5B · Deprecated

Use the new LFM2.5-Audio-1.5B checkpoint instead.

Examples

Explore practical implementations using audio models:

Real-Time Audio Transcription CLI

Transcribe audio files locally in real-time using LFM2-Audio-1.5B with llama.cpp for 100% private, on-device processing.

Platform: Desktop · Uses: LFM2-Audio-1.5B

Browser Audio Demo with WebGPU

Run ASR, TTS, and interleaved conversations entirely in-browser using LFM2.5-Audio-1.5B with WebGPU acceleration.

Platform: Web · Uses: LFM2.5-Audio-1.5B