Fish Speech

Fish Speech is an open-source text-to-speech (TTS) platform developed by Fish Audio that delivers real-time voice synthesis with fine-grained emotion control and zero-shot voice cloning. Built on a dual autoregressive architecture (VQGAN + Llama), it supports over 13 languages including English, Mandarin, Japanese, Korean, French, German, Arabic, and Spanish, making it one of the most multilingual open-source TTS solutions available as of early 2026.

The platform allows users to clone a voice from as little as 10–15 seconds of reference audio, producing natural-sounding speech that preserves the tone, cadence, and stylistic qualities of the source. Emotion control is achieved through prompt engineering and reference audio selection, enabling users to generate speech with specific emotional inflections such as happiness, sadness, anger, or calm without retraining the model.

Fish Speech operates with inference latency under 150 milliseconds on consumer-grade GPUs (RTX 3060 and above), enabling real-time or near-real-time voice generation suitable for interactive applications like chatbots, virtual assistants, and live content creation. The model weights are released under the Apache 2.0 license for the base model, with a commercial-friendly CC-BY-NC-SA 4.0 license for certain fine-tuned checkpoints.

The Fish Audio cloud platform (fish.audio) provides a hosted API with managed infrastructure, removing the need for local GPU resources. The API supports streaming audio output, batch processing, and SSML-like markup for controlling pacing, pauses, and emphasis. As of version 1.5 (released Q1 2026), the model achieves a Mean Opinion Score (MOS) of approximately 4.1 out of 5.0 on standard speech naturalness benchmarks, competitive with proprietary solutions that cost significantly more.

Fish Speech has accumulated over 15,000 GitHub stars and an active open-source community contributing voice packs, language fine-tunes, and integration plugins for platforms such as Discord, OBS Studio, and Unity. The platform processes over 2 million API requests per month through its hosted service, serving creators, developers, and enterprises across gaming, audiobook production, accessibility tools, and customer service automation.