Liquid AI releases LFM2-Audio-1.5B speech model
Liquid AI has released LFM2-Audio-1.5B, a compact end-to-end model that handles speech recognition, text-to-speech and real-time voice chat in one system, aimed at low-latency use on devices.
Liquid AI announced LFM2-Audio-1.5B, which adds audio to its LFM2 family. It is built on the LFM2-1.2B language model and can take text or audio as input and produce either one as output.
Input audio and output audio are represented differently. Incoming audio skips a tokenizer: the raw waveform is cut into segments of about 80 ms that are projected straight into the backbone's embedding space as continuous features. For output, the model predicts discrete audio tokens the same way it predicts text tokens, generating up to eight per step, and then decodes them into a waveform. Liquid says this lets it train the model as a single next-token predictor while avoiding artifacts from discretizing the input.
On VoiceBench, a set of nine audio-interaction benchmarks, the model scored 56.78 overall. That beats Moshi (7B, 29.51) and Mini-Omni2 (0.6B, 33.49) but trails Qwen2.5-Omni-3B (63.57), which Liquid lists at 5B parameters. Its average ASR word error rate was 7.24, compared with 7.93 for Whisper-large-v3, 7.15 for Qwen2.5-Omni-3B and 6.47 for ElevenLabs' closed Scribe v1.
Liquid reports average end-to-end latency under 100 ms, measured from a 4-second spoken input to the first audible output. The post does not say what hardware was used.
The model is available on Hugging Face, GitHub and Liquid's playground. It comes with a Python inference package and a reference speech-to-speech chat app. The post does not mention licensing terms.