FunCodec

FunCodec is a fundamental neural audio coding framework developed by the Alibaba DAMO Academy (FunAudioLLM team). Designed to support the burgeoning Large Audio Model (LAM) ecosystem, FunCodec employs Residual Vector Quantization (RVQ) to map continuous audio waveforms into discrete tokens. Unlike traditional codecs such as Opus or MP3, FunCodec is optimized for both acoustic reconstruction and semantic extraction, making it a critical bridge for multimodal Large Language Models that process speech directly. As of 2026, it remains a cornerstone for developers building low-latency speech-to-speech translation systems and high-compression audio storage solutions. The architecture supports multiple backbones, including CNNs and Transformers, and offers frequency-domain modeling to enhance perceptual quality. It is uniquely positioned in the market as an alternative to Meta's EnCodec and Google's SoundStream, offering superior performance in Chinese speech processing and seamless integration with the FunASR and CosyVoice ecosystems. Its ability to maintain high fidelity at extremely low bitrates (sub-1kbps) makes it ideal for edge computing and bandwidth-constrained environments.