LFM2-Audio-1.5

LFM2-Audio-1.5 brings ASR, TTS, and audio generation into a single model so teams can move from text or reference audio to finished sound without stitching tools together. It listens and speaks in the same session, preserving context across turns, and can follow detailed direction about voice, style, tempo, and atmosphere while keeping timing predictable for live agents or interactive media. On understanding tasks it produces clean transcripts with punctuation, speaker attribution when needed, and reliable timecodes that line up in an editor. On generation it renders expressive speech, creates music or sound effects that match the brief, and performs localized edits—re-voicing lines, extending ambience, inpainting gaps, or removing distractions—without breaking continuity or loudness. The model adapts to multiple languages and accents, accepts style references to maintain brand or character identity, and can perform consented voice matching under policy controls. Designed for production, it streams partial results, emits structured JSON alongside audio for automation, and fits into existing pipelines with quantization options for edge or high-throughput serving.