03/04/2026
ParrotSpeech TTS: Hybrid Autoregressive and Flow-Matching Text-to-Speech
ParrotSpeech TTS is a sophisticated multilingual text-to-speech system capable of reproducing high-quality speech with just three seconds of reference audio. The model utilizes a hybrid architecture that generates automatically recursive semantic tokens while employing stream-matching transducers to produce detailed audio tokens. This process is supported by the ParrotSpeech Codec, a specialized encoder that efficiently compresses audio into an arithmetic representation by combining vector and scalar quantization. Human evaluations show that the system significantly outperforms commercial competitors like ElevenLabs in terms of naturalness and speaker similarity across nine different languages. To further enhance performance, the developers employed Direct Priority Optimization (DPO) to fine-tune both discrete and continuous generation stages. Furthermore, the software framework is optimized for low-latency streaming via vLLM-Omni, utilizing CUDA graph acceleration capabilities to ensure rapid, real-time speech synthesis.
