SpeechTokenizer: Unified Speech Tokenizer for Speech Large Language Models

Current speech large language models build upon discrete speech representations, which can be categorized into semantic tokens and acoustic tokens. However, existing speech tokens are not specifically designed for speech language modeling. To assess the suitability of speech tokens for building speech language models, we established the first benchmark, SLMTokBench. Our results indicate that neither semantic nor acoustic tokens are ideal for this purpose. Therefore, we propose SpeechTokenizer, a unified speech tokenizer for speech large language models. SpeechTokenizer adopts the Encoder-Decoder architecture with residual vector quantization (RVQ). Furthermore, We construct a Unified Speech Language Model (USLM) leveraging SpeechTokenizer. Experiments show that SpeechTokenizer performs comparably to EnCodec in speech reconstruction and demonstrates strong performance on the SLMTokBench benchmark. Also, USLM outperforms VALL-E in zero-shot Text-to-Speech tasks. Large language models (OpenAI, 2023; Touvron et al., 2023) have demonstrated remarkable performance on various natural language processing tasks. This has inspired numerous works to build speech language models (Borsos et al., 2022), which have achieved significant breakthroughs across various speech processing tasks (Wang et al., 2023; Zhang et al., 2023; Rubenstein et al., 2023; Dong et al., 2023). A key commonality among these works is the utilization of discrete speech representations.