On July 28, 2026, Fish Audio announced a $52M seed round co-led by Coreline Ventures and Capital Today, with participation from 359 Capital, Parable, Play Time, Alphalist Partners, Bayhouse Ventures, Carya Venture Partners, HF0, and 645 Ventures.

The round was Fish Audio's first institutional financing, bringing total funding to $52M. CEO Rissa Cao founded the company out of an open-source voice-generation project started by former Nvidia researcher Shijia Liao.

Fish Audio turns text into speech with controls for emotion, pacing, and delivery. Users paste a script, select from more than 2 million community-uploaded voices or clone a voice from a 10-to-30-second recording, then add inline natural-language instructions that can change the performance mid-sentence.

Rather than relying on global speed or pitch sliders, Fish Audio supports over 15,000 natural-language vocal directions. These include tags such as [whisper], [angry], and [laughing], as well as free-form descriptions such as [professional broadcast tone].

Instant cloning extracts a speaker's identity from a short audio sample and generates speech in that voice across more than 80 languages, including when the source recording uses a different language. Professional cloning provides verified, higher-fidelity voice records for commercial use by voice actors, brands, and studios.

For developers, Fish Audio offers REST and WebSocket APIs with streaming, official Python and JavaScript SDKs, and controls for latency-quality trade-offs. In optimized conditions, the S2.1 Pro model targets sub-100-millisecond time-to-first-audio for real-time voice-agent applications.

The platform also provides speech-to-text transcription across 80-plus languages, text-prompted synthetic voice generation through Voice Design, long-form multi-speaker production through Story Studio, and a voice-agent builder with system prompts, knowledge bases, and session management. Users can transcribe existing audio, edit it as text, translate it, and regenerate it with the original or a new voice.

Fish Audio sells through three channels: self-serve creator subscriptions, pay-as-you-go developer APIs, and custom enterprise contracts. Creator plans range from free (approximately 7 minutes of generation per month) to $999/month for Max (up to 6,250 minutes, 10 seats, and professional voice slots). These prepaid generation bundles generate recurring revenue while capping compute consumption per customer.

The API is billed separately at $15 per million UTF-8 input bytes for TTS and $0.36 per audio hour for speech recognition, with no monthly minimums. Concurrency scales with cumulative prepaid spend, from five concurrent requests below $100 to 50 at $1,000 and custom limits for enterprise. Requiring customers to prepay to unlock capacity before consuming it also provides working capital.

Enterprise contracts combine volume pricing, reserved capacity, SLAs, zero-data-retention options, and on-premise deployment. Fish Audio reached more than $6M in enterprise annualized revenue before building a dedicated sales team, with early enterprise sales handled by the founders and driven by product adoption.

The open-weight Fish Speech model, which accumulated over 31,000 GitHub stars, acts as a distribution funnel. Developers discover Fish through the open repository and experiment with self-hosted inference, then can convert to the hosted API or enterprise tier for the latest S2.1 Pro model, optimized infrastructure, or production SLAs. The model's research license restricts direct commercial competition using the open weights, protecting the paid offering while retaining community-based acquisition.

Fish Audio operates its own GPU infrastructure and has built custom FP8 CUDA kernels and scheduling systems that reportedly reduced the number of H200 GPUs needed for a given workload by 4x and pushed utilization above 90%. The lower cost per generated second enables pricing that includes a temporary free S2.1 Pro API through November 2026 as a product-led acquisition program. Fish Audio also pays a 20% affiliate commission for up to 12 months on qualifying subscription sales, using YouTube tutorials and social demonstrations to reach its creator audience.

Fish Audio competes across creator voice generation, low-latency speech infrastructure for voice agents, and enterprise audio platforms. Competition is converging around bundled stacks rather than standalone TTS endpoints.

ElevenLabs is the most direct competitor, with an estimated $500M in annualized revenue as of 2026 and products spanning cloning, dubbing, sound effects, music, speech-to-text, and conversational agents. ElevenLabs can cross-subsidize TTS through adjacent products and offers enterprise controls including SSO, HIPAA options, and audit tooling. Fish Audio competes on price, expressiveness, and open-model distribution, while ElevenLabs can win accounts where procurement simplicity and workflow breadth outweigh model-level quality differences.

MiniMax is a close technical rival in emotional speech and multilingual generation, offering turbo and HD models, rapid cloning, and aggressive pricing. Its multimodal portfolio spans text, video, and music generation, particularly for developers already using its platform. Fish Audio differentiates through its larger public voice community, more extensive natural-language control, and open-source developer reputation outside China.

Cartesia has broadened from a latency-focused TTS engine into an agent stack with speech recognition, managed agents, and telephony. Its $191M in total funding supports product expansion. Developers purchasing Cartesia's managed-agent product no longer need to select Fish Audio separately as the TTS provider, making Cartesia a direct threat to Fish Audio's role as a component supplier.

Deepgram entered TTS from a speech-to-text installed base and now offers a unified Voice Agent API combining transcription, synthesis, and orchestration at $4.50 per hour. Existing Deepgram STT customers can add TTS without introducing another vendor. Its domain-specific pronunciation and self-hosting options target regulated contact centers where compliance matters more than expressive range.

Smallest.ai targets the price-sensitive real-time market with sub-100ms latency claims and rates around $0.0175 per 1,000 characters, with particular strength in South Asian languages and code-switching. Its developing speech-to-speech stack could bypass standalone TTS selection entirely.

OpenAI's Realtime API combines speech recognition, reasoning, and speech generation in a single model interaction, removing the need to compose separate STT, LLM, and TTS providers. This presents a structural threat to modular TTS vendors by reducing network hops and engineering complexity, even when its per-minute economics are less favorable than a dedicated low-cost TTS stack.

Agent orchestration platforms such as Retell AI, Vapi, and Bland AI sit between voice-model vendors and end customers. Fish Audio counts Retell, LiveKit, and Telnyx among its integration partners, but these platforms can source from multiple models, dynamically route traffic based on price or latency, and negotiate wholesale rates. This lowers switching costs and compresses Fish Audio's pricing power.

Google Cloud, Microsoft Azure, and AWS present a similar consolidation risk through procurement bundling, as customers already using a hyperscaler can add speech without introducing another vendor.

Fish Audio's expansion strategy extends from a TTS API to a broader audio intelligence platform, while widening its customer base across creators, developers, enterprises, and geographies.

Fish Audio is developing speech-to-speech and audio-understanding models that would let it capture more of each voice interaction, rather than supplying only the final synthesis layer. Direct speech-to-speech preserves emotion, cadence, and interruptions lost in the traditional ASR-to-LLM-to-TTS pipeline, capabilities required by conversational agents.

The company already distributes through voice-agent and communications platforms including Retell, LiveKit, Telnyx, and Sanas. Deeper integrations with turn detection, interruption handling, telephony observability, and call analytics could increase switching costs and shift Fish Audio from a metered synthesis supplier to a production control plane for voice applications. Sacra estimated Retell AI at $60M in annualized revenue in April 2026, indicating the scale of the voice-agent infrastructure market available to Fish Audio.

Fish Audio's roots in VTuber, anime, gaming, and indie creator communities provide a path into multilingual dubbing, audiobook narration, game dialogue, podcasts, and advertisements. Its fine-grained emotional and paralinguistic controls are more relevant in performative media than in utility TTS, giving it an entry point into workflows where directing a virtual performance matters more than reading text aloud.

Potential additions include automated translation and dubbing with timing correction, speech-to-speech performance transfer across languages, and integrations with animation and lip-sync systems. Its existing partnership with HeyGen shows how Fish Audio can operate as an embedded voice engine within video and avatar platforms, rather than acquiring every end user directly.

Professional voice cloning could develop into a two-sided licensing marketplace connecting voice actors, brands, and developers. Fish Audio introduced professional cloning with verified ownership workflows, providing the basis for consent-controlled voice registries in which rights holders authorize specific languages, territories, and media types and receive compensation when their voices are used.

Enterprise expansion covers healthcare, financial services, government, and contact-center markets, where voice data can contain protected information. Fish Audio advertises on-premise deployment, zero-data-retention, and HIPAA-oriented support. Bundling ASR and TTS into per-agent or per-conversation enterprise pricing, rather than selling separate API calls, could increase account values and switching costs. Support for 83 languages provides a technical base for geographic expansion, though adoption also requires regional accents, local pronunciation, in-country compliance, and dedicated inference capacity beyond the stated language count.

Voice consent and impersonation: Fish Audio's instant cloning from short audio samples and library of over 2 million community-uploaded voices expose the company to unauthorized celebrity and creator clones, publicity-right violations, and fraud, while its automated DMCA process removes disputed voices within minutes but remains reactive, allowing unauthorized clones to persist until the affected person discovers and reports them.

Model commoditization and price compression: As ElevenLabs, Cartesia, Deepgram, OpenAI, MiniMax, Smallest.ai, and hyperscalers converge on high-quality multilingual speech generation, standalone TTS risks becoming a commodity, with orchestration platforms such as Retell AI, Vapi, and Bland AI dynamically routing traffic among interchangeable providers based on price and latency, compressing margins and reducing switching costs for Fish Audio's developer and enterprise customers.

Platform disintermediation: OpenAI's Realtime API and other native speech-to-speech models can collapse the ASR-LLM-TTS pipeline into a single model call and bypass modular TTS providers such as Fish Audio, while the agent platforms and communications companies that distribute Fish Audio's technology could build proprietary voice capabilities or negotiate wholesale rates that erode Fish Audio's pricing power and strategic visibility.