AmphionTeam/FlexiCodec

Implements dynamic frame rate encoding that adaptively merges semantically similar frames (3–12.5Hz) using ASR-assisted dual-stream encoding and Transformer bottlenecks to preserve semantic information at very low frame rates. Decouples semantic and acoustic codes for efficient speech language model integration, with inference-time controllable compression via `merging_threshold`. Includes AR+NAR TTS pipeline and Voicebox NAR decoder for speech synthesis from FlexiCodec tokens.