1. Tutorial Introduction

Kyutai TTS 1.6B (en-fr) is a large-scale English-French bilingual speech model released by the Kyutai team on October 15, 2024. In streaming TTS benchmarks, this model outperforms traditional offline TTS by 751 TP3T and 421 TP3T in the "real-time output of long texts" and "bilingual prosodic naturalness" categories, respectively. It also achieved state-of-the-art performance in TTS benchmarks such as Moshi Benchmark and Audio-Language Alignment Dataset. Furthermore, the model demonstrates features rarely seen in previous systems, including input-output streaming generation, zero-shot switching between English and French, speech selection based on pre-computed embeddings, and fast inference with dynamically adjusted audio token counts. Related paper results are available. Streaming Sequence-to-Sequence Learning with Delayed Streams Modeling.

This tutorial uses a single RTX 4090 graphics card. Only English and French are supported.

2. Project Examples

standard-tts

streaming-tts

3. Operation steps

1. After starting the container, click the API address to enter the Web interface

2. Usage steps

If "Bad Gateway" is displayed, it means the model is initializing. Please wait approximately 2-3 minutes and then refresh the page. When using the Safari browser, audio may not play directly and needs to be downloaded first.

Citation Information

@techreport{kyutai2025streaming,
      title={Streaming Sequence-to-Sequence Learning with Delayed Streams Modeling}, 
      author={Neil Zeghidour and Eugene Kharitonov and Manu Orsini and Václav Volhejn and Gabriel de Marmiesse and Edouard Grave and Patrick Pérez and Laurent Mazaré and Alexandre Défossez},
      year={2025},
      eprint={2509.08753},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2509.08753}, 
}