Fish Audio Ships Dual-AR Speech Model Alongside $52M Seed

Palo Alto-based startup Fish Audio has closed a $52 million seed round and reached $21 million in annual recurring revenue just one year after its launch. The July 28 announcement coincides with the release of Fish Speech S2.1 Pro, the company’s new flagship speech generation model. The round was co-led by Coreline Ventures and Capital Today, validating a platform that now serves over 8 million users across its open-source and hosted deployments.

Dual-AR Architecture and Performance

The Fish Speech S2 family moves away from standard diffusion or continuous auto-regressive pipelines in favor of a Dual-Autoregressive (Dual-AR) architecture. This system separates the generation process into two specialized components. A 4-billion-parameter “Slow AR” model predicts discrete semantic tokens, while a 400-million-parameter “Fast AR” model renders the final acoustic details.

This structural split allows the inference engine, built on SGLang, to process requests rapidly. The system achieves a Real-Time Factor (RTF) of 0.195 and a Time-to-First-Audio (TTFA) of under 100ms. The models were trained on more than 10 million hours of audio data.

The resulting S2.1 Pro model supports 83 languages and introduces granular, word-level emotion control. Developers can inject over 15,000 natural-language tags, such as [whisper], [laughing], or [angry], directly into prompts to manipulate tone mid-sentence. In blind listening tests, the company notes that 66% to 67% of evaluators preferred S2.1 Pro over leading competitors.

While Fish Audio maintains three open-source models in a GitHub repository with over 31,000 stars, S2.1 Pro is closed-source. The company is offering it for free via API through the end of August 2026.

Enterprise Infrastructure and Safety

Fish Audio is positioning itself as the core infrastructure layer for companies building real-time voice agents and localization tools. The current enterprise roster includes OpenAI, HeyGen, Sanas, LiveKit, Plaud, and Retell AI.

To compete with established providers like ElevenLabs, the platform supports HIPAA-compliant configurations, on-premises deployment, and zero-data-retention environments required by regulated industries.

Operating at this scale requires automated governance. To mitigate malicious voice cloning, Fish Audio deployed an automated takedown system that removes a cloned voice within three minutes of receiving a verified request from the original voice owner.

Roadmap and Market Expansion

With the new capital, the 22-person team is expanding its focus beyond text-to-speech. The product roadmap targets the development of voice-native Large Language Models (LLMs) and real-time speech-to-speech translation pipelines. The company will also build a dedicated enterprise sales division to scale its commercial footprint.

If you manage high-volume speech generation workloads, the temporary free tier for the S2.1 Pro API provides a practical window to evaluate its 100ms latency and emotion tagging capabilities against your current vendor. The Dual-AR architecture offers a distinct performance profile that may reduce bottlenecks in conversational applications.