VoxCPM2: The ambition of this 2B model is not just “another TTS”

Normal people think that the end point of speech synthesis is to make the model bigger and feed more data. XTTS, CosyVoice, F5-TTS, everyone follows this path: the text is entered, cut into discrete tokens, the language model predicts the token sequence, and the vocoder restores the waveform. This process has been running for three or four years, and everyone is busy comparing whose MOS score is a few tenths higher and who has a few milliseconds less delay. VoxCPM2 did not do this, it directly deleted the Tokenizer from the pipeline. This is not an academic gimmick. Throwing away the Tokenizer means that the model does not need to compress a piece of speech into 1024 discrete symbols and then restore it, but directly models the sound in a continuous space, skipping the bottleneck of quantization and compression. What's the result?