Most people’s mental model of text-to-speech is frozen somewhere around 2018. Robotic delivery, limited voice selection, obvious artifacts on anything longer than a sentence, and a pricing model that made high-volume use impractical for anyone without a dedicated audio budget. That picture is significantly out of date, and the gap between what people assume TTS can do and what it actually does right now is wide enough to affect real decisions.
The changes aren’t incremental. Quality, latency, pricing, and language coverage have all moved at the same time — the rare combination that shifts a technology from a specialized vendor decision into something you can actually build around. Current-generation voice models are clearing benchmarks that earlier models failed by significant margins, and they’re doing it at price points that make high-volume generation practical for the first time.
AI text to speech is the entry point for most users — paste a script, hear what comes back, adjust and regenerate. The basic loop is now fast enough that it feels more like editing text than commissioning a production. What’s changed is what’s available inside that loop: natural-language control over emotional delivery, voice cloning from a short sample, 80-plus languages from a single model, and audio that comes back fast enough to use in live, conversational applications.
The Quality Jump Is Real and Measurable
The hardest claim to evaluate in any AI category is quality, because vendors all say the same thing. Published benchmark data is more useful than marketing copy, and the current numbers are genuinely surprising.
Fish Audio ran a blind A/B test on real production traffic — over 5,000 preference pairs, where the “winner” was whichever version a listener actually downloaded after hearing both at least twice. Its S2 Pro model beat ElevenLabs V3 60% to 40% in direct head-to-head comparison. On the Audio Turing Test — a benchmark specifically designed to test whether listeners can distinguish synthetic speech from a human voice — the same model scored 0.515, which is above the threshold where humans can reliably tell the difference. On EmergentTTS-Eval, it posted an 81.88% win rate against a GPT-4o-mini-TTS baseline, including 91.61% on paralinguistic delivery, the metric that measures how well the model executes emotional instruction rather than just reading text cleanly.
The current-generation S2.1 Pro model has since posted a 61% win rate against S2 Pro in the same head-to-head format. The models keep improving generation over generation.
Latency: Fast Enough for Real-Time
The latency story is the one that matters most for what you can actually build with TTS. Earlier generation models were batch-first — you sent a request and got audio back seconds later, which was fine for pre-rendered content but ruled out anything interactive.
Current leading models post time-to-first-audio in the 70–100ms range. That number matters because 200–300ms is roughly the threshold where a response pause becomes perceptible as a delay in conversation. Under 100ms, voice generation can run inline in a product flow — conversational IVR, live voice assistants, interactive avatar pipelines — without the “thinking pause” that makes automated interactions feel obviously robotic. Fish Audio’s S2.1 Pro runs in that range under normal load.
Emotion Control: Not a Dropdown Anymore
Most older TTS systems handled emotion through fixed preset moods — a dropdown menu with options like “happy,” “sad,” “neutral.” The limitation is obvious: a small number of fixed states can’t cover the actual range of delivery a real script requires.
Current systems use open-domain natural-language tags embedded directly in the script. Fish Audio’s approach lets you write instructions like [reassuring] or [the calm, measured tone of someone who has done this a thousand times] directly into the text, at the word level. The model interprets these as natural-language direction rather than mapping them to a fixed category. A single sentence can shift tone mid-phrase. For any use case where delivery carries meaning — customer service, training content, brand narration — that flexibility is a meaningful capability upgrade over a preset system.
AI Voice Cloning: 15 Seconds to a Reusable Voice
AI voice cloning is now a standard feature on most serious platforms, not a premium add-on. Fish Audio can generate a consistent, reusable voice model from a reference sample as short as 15 seconds. Once created, the cloned voice is stored as a callable parameter — so every generation that uses it sounds like the same person, regardless of when it was produced or what script it’s reading.
The practical implications: a content creator can define a voice once and maintain it indefinitely. A brand can establish a spokesperson voice and deploy it across every campaign without rebooking. A developer building a voice assistant can give it a consistent identity from a single short recording.
One important note: AI voice cloning requires a paid commercial plan for production use. Free tiers on most platforms, including Fish Audio, are restricted to personal, non-commercial use.
Language Coverage: One Model, 83 Languages
Fish Audio’s S2.1 Pro covers 83 languages from a single endpoint. That’s not just a big number — the architecture matters. A single-endpoint multilingual model means the same API call handles every language your application needs, rather than routing to different underlying systems with different quality levels per language.
The practical difference: localization that used to mean a separate vendor per market now means a language parameter in the same request. For anyone building multilingual applications or producing content across multiple markets, that compression in the workflow is significant.
What It Actually Costs
API pricing is usage-based — Fish Audio charges $15 per million characters generated, with no subscription requirement. You pay for what you generate. For reference, independently published pricing from competitors like ElevenLabs has run closer to $165 per million characters, though pricing in this category moves and should always be checked against current rate cards.
Plan-based pricing for non-API use starts with a free personal tier and a Plus plan at $11/month that includes commercial use rights. The free tier covers limited generation for personal use only — it’s not licensed for commercial content.
Speech Recognition: The Other Side of the Stack
Modern voice platforms aren’t just generation. Fish Audio’s ASR (automatic speech recognition) runs at $0.36 per audio hour and returns multi-speaker labeled output with timestamps automatically. For anyone handling call transcription, meeting notes, podcast transcripts, or content indexing, that price point makes programmatic transcription practical at almost any scale.
The Community Voice Library
Fish Audio maintains a community voice library with over 2,000,000 voices contributed by users. This is a fast way to explore different voice styles without commissioning a custom clone — though community voices should be treated as platform-licensed assets rather than cleared-for-any-purpose content. Check per-voice terms before deploying in production.
Open-Weights for Those Who Need Control
Fish Audio releases model weights, fine-tuning code, and inference engines publicly, making self-hosted deployment an option. The correct term is open-weights, not open-source — the weights are downloadable and runnable, but commercial use still requires a paid license. For developers or organizations with data residency requirements or regulated-industry constraints, this is worth knowing exists. For everyone else, the managed API is the simpler path.
Where to Start
The lowest-friction way to calibrate whether current TTS meets your needs: run one real script — at full length, not a short demo clip — through a free account and listen critically. The free tier is personal use only, so commercial evaluation should move to a paid plan, but the technical capability is identical. What you’ll likely find is that the mental model built on 2018-era TTS doesn’t match what you hear back — and that the gap is larger than expected.
