Fish Audio Opens S2.1 Pro to Developers With Free Unlimited API Access

Fish Audio's S2.1 Pro , the same model powering its paid tier , is now available to any developer for free via API. S2.1 Pro is now available as a free text-to-speech API with unlimited access under Fair Use. The move flips the standard industry script, where production-quality voice has always sat behind a paywall.

The paywall problem it's solving

The models that actually sound good cost money. ElevenLabs' free tier gives you roughly 6–10 minutes of audio before the paywall kicks in. OpenAI TTS has no free tier at all. Google's latest Gemini TTS models have zero free usage , you pay from the first token.

The AI voice generator market is growing at nearly 20% annually, but the tooling to build voice-enabled products has stayed behind a paywall. You can't properly evaluate a model on 10,000 credits. Fish Audio is betting that removing the barrier entirely will accelerate adoption , and that their new inference stack makes it economically viable.

What S2.1 Pro actually is

S2.1 Pro is Fish Audio's current state-of-the-art voice model, designed for production-grade AI voice generation, with particular strengths in low-latency streaming, multilingual TTS, and voice cloning. It's an improved version of S2 Pro, which Fish Audio released with open weights earlier this year.

The headline improvements over the previous generation:

• 61% win rate against S2 Pro in head-to-head listening evaluations
• ~70ms Time-to-First-Audio (TTFA) at single request , down from ~100ms in the prior generation
• 2x+ throughput improvement under high-concurrency load
• 83 languages supported, up from 80+ in S2-Pro

Under the hood: why it sounds this good

S2 Pro was trained on over 10 million hours of audio data covering more than 80 languages, combining a Dual-Autoregressive (Dual-AR) architecture with reinforcement learning alignment to generate speech that is exceptionally natural, realistic, and emotionally rich.

The Dual-AR architecture is the key structural innovation. Instead of one model doing everything, there are two working in tandem:

• Slow AR (4B parameters): operates along the time axis and predicts the primary semantic codebook , essentially handling what is being said.
• Fast AR (400M parameters): generates the remaining 9 residual codebooks at each time step, reconstructing fine-grained acoustic detail , handling how it sounds.

This asymmetric design keeps inference efficient while preserving audio fidelity. Because the Dual-AR architecture is structurally isomorphic to standard autoregressive LLMs, it inherits all LLM-native serving optimizations. That means techniques like continuous batching and paged KV cache , normally reserved for language model serving , apply directly to voice generation.

On the training side, Fish Audio used GRPO (Group Relative Policy Optimization , a reinforcement learning technique that scores model outputs against each other to improve quality) for post-training alignment. A core architectural decision is that the same models used to filter and annotate training data are directly reused as reward models during RL. The rich-transcription ASR model generates caption-augmented transcripts during data curation, then provides the instruction-following reward by re-transcribing generated audio and comparing it against the original prompt. This closed loop eliminates the distribution mismatch that plagues most TTS training pipelines.

The tag system: directing speech like a screenwriter

S2.1 Pro treats bracket tags as standard text rather than dedicated control tokens. Through training on massive datasets, the model learned implicit mappings between natural language descriptions and acoustic variations , meaning you're not limited to a predefined set of tags. You can write anything descriptive and the model will interpret it.

Some examples of what this looks like in practice:

# Inline emotion tags placed anywhere in the text
text = """[professional broadcast tone] Welcome to today's report.
[whispers] But between us , [laughing nervously] I probably shouldn't say this.
[angry] This is completely unacceptable!"""

On EmergentTTS-Eval, S2 achieves particularly strong results in paralinguistics (91.61% win rate), questions (84.41%), and syntactic complexity (83.39%). On the Audio Turing Test , a benchmark that asks whether generated speech sounds human , S2's score of 0.515 surpasses Seed-TTS (0.417) by 24% and MiniMax-Speech (0.387) by 33%.

Voice cloning, included

S2 supports accurate voice cloning using a short reference sample (typically 10–30 seconds). The model captures timbre, speaking style, and emotional tendencies, producing realistic and consistent cloned voices without additional fine-tuning. This works across all 83 supported languages, making it especially useful for content localization where consistent speaker identity is critical. Voice cloning is included in the free tier.

Where it falls short

S2.1 Pro isn't perfect across the board:

• MiniMax-Speech and ElevenLabs still maintain an advantage in certain low-resource languages , typically those with under 1,000 hours of training data.
• Some users report that ElevenLabs retains an edge in pure naturalness for certain highly specific English voices. The gap is narrowing, but it still exists for the most demanding use cases.
• The free tier carries no SLA or latency guarantee , it's built for experimentation and prototyping, not contractual uptime.

Why the free tier is economically sustainable

This is the part that makes the announcement more than a marketing move. Fish Audio rebuilt the inference stack from the ground up. They developed fish-scales-ops, a production-grade FP8 GEMM and FlashAttention library targeting NVIDIA Hopper (H100/H200) and Blackwell architectures. On the decode shapes that matter for voice AI serving, their MXFP8 path outperforms the torch.compile-fused cuBLAS reference by 2.1–4.3x. More throughput per GPU means more requests served per dollar , which is what makes unlimited free access viable rather than a loss-leader stunt.

How to get started in two minutes

If you're already calling the Fish Audio API, the migration is a single line change. Set model: "s2.1-pro-free" in the headers , that's the only change from any other Fish Audio API call.

import httpx
body = {
    "text": "Hello, world!",
    "reference_id": "your_model_id",
    "format": "mp3",
}
with httpx.Client() as client:
    res = client.post(
        "https://api.fish.audio/v1/tts",
        headers={
            "Authorization": "Bearer YOUR_API_KEY",
            "Content-Type": "application/json",
            "model": "s2.1-pro-free",   # <-- this is the only change
        },
        json=body,
    )
with open("output.mp3", "wb") as f:
    f.write(res.content)

Get an API key at fish.audio/app/api-keys. Full reference in the official API docs.

What you can build with it

The free tier is intentionally broad on use cases. The strongest fits:

• Voice agents: Real-time conversational AI lives and dies by latency. At ~90ms TTFA for standard calls, S2.1 Pro is fast enough for natural turn-taking dialogue.
• Audiobooks and long-form narration: 83-language support and natural prosody make S2.1 Pro well-suited for audiobook production and long-form speech synthesis.
• Multilingual apps: 83-language coverage with a single consistent AI voice API is a meaningful simplification over alternatives that require separate model endpoints per language or charge premium rates for non-English speech synthesis.
• Game NPC dialogue: Game audio pipelines benefit from high throughput and predictable cost per request. Unlimited free usage makes it practical to generate large dialogue libraries and iterate freely during development before committing to a production budget.

The fine print

Free access is available through the end of July 2026, with advance notice before any changes. There's no SLA, no latency guarantee, and requests may be used to improve model quality. Products generating more than $1M ARR should contact Fish Audio before using S2.1 Pro Free. For production workloads that need contractual guarantees, paid plans remain available. The open-source model weights and inference stack on GitHub are also an option for teams who want to self-host entirely.