Best Real-Time Text-to-Speech APIs for Game Characters and AI Companions (2026)

ElevenLabs is the strongest all-round choice for voiced characters because it covers both halves of the job: designing a voice that sounds like a person, and serving it fast enough for live dialogue. For runtime lines you pick a latency-tuned model. Flash lists about 75 ms of model latency across 32 languages, and the newer v4 Turbo is positioned as the model for real-time agents at about 100 ms with v4-level expressiveness. For cutscenes, trailers and barks you can pre-generate, v3 and v4 add audio tags for laughs, whispers and emotional shifts.

The tooling around the model is where it pulls ahead for character work. Voice Design creates a new voice from a text description ("gravelly old dwarf, tired, kind"), instant and professional cloning lets you license a real actor's voice, and every voice gets a stable ID you can reuse for months. A WebSocket input-streaming endpoint accepts LLM text as it is generated, so audio can start before the reply is finished.

The trade-off is cost at companion-app scale. API list prices run $0.04 per 1,000 characters for Flash and Turbo-class models and $0.08 for v3, roughly 2.5 to 5 times the per-character rates of Inworld AI or Fish Audio. Subscription plans from $6 to $990 a month bundle credits and raise concurrency, and commercial use requires a paid plan. For a studio voicing a few hero characters that cost is easy to justify. For a free-to-play companion with heavy daily listeners, model the bill before you commit.

Inworld AI began as a character engine for games, and its current voice stack still reads like it was built by people who have shipped NPCs. Realtime TTS-2 is the expressive model, with a vendor-stated P99 time to first byte under 100 ms. TTS-2 Flash trades a little expressiveness for 25 ms P99 and a lower price. Both stream over WebSocket and take natural-language steering inline, such as [Speak sadly] or [Speak softly], plus non-verbal cues like [laugh], [sigh] and [breathe], which is exactly how a dialogue system wants to direct a performance per line.

The economics are the real argument. On demand, TTS-2 costs $25 and Flash $15 per million characters. Paid plans lower that, down to $12.50 and $7 on Growth, and enterprise pricing goes as low as $5 per million for TTS-2. Concurrency is generous: 50 concurrent requests on the $100 Builder plan, and Inworld estimates each request slot serves at least four user sessions. For a companion app or a multiplayer game with many talking NPCs, that headroom matters as much as voice quality.

The rest of the platform fits a character pipeline too: speech-to-text that returns emotion, pitch and vocal-style signals from the player's voice, a speech-to-speech Realtime API with barge-in, and an LLM router that can send cheap combat barks to a fast open model and story-critical scenes to a stronger one. Engine integration is through standard REST, WebSocket and WebRTC APIs, so you write the Unity or Unreal glue yourself.

Cartesia is the pick when the conversation has to feel instant. Its Sonic models run on a state-space architecture rather than a standard transformer, and Sonic-3.6 lists first audio in under 90 ms across 44 languages from a single model. Independent reviews put real-world turn latency higher once network time is included, which is true of every vendor here, but Cartesia sits consistently near the front of public latency benchmarks.

For characters, the most useful feature is contexts. You open a context over WebSocket and keep pushing text chunks from your LLM into it. Sonic treats them as one continuous utterance, so prosody doesn't reset at every chunk boundary, which is the robotic stitching effect that gives away a lot of streamed NPC dialogue. Sonic also reads emotional subtext from the transcript by default, accepts non-verbal cues like [laughter] inline, and offers volume, speed and emotion controls. Custom pronunciation dictionaries handle invented place names and spell names, which fantasy games are full of.

Pricing is credit-based, with one character equal to one credit, and starts low: $5 a month buys 100,000 credits, a commercial license and instant cloning from about 10 seconds of audio. The constraint is concurrency on lower plans: 2 to 5 concurrent TTS requests below the $299 Scale plan, and 15 on Scale. That is fine for a single-player game where one conversation happens at a time, and tight for a companion app at launch.

Fish Audio lists AI companions among its target use cases, and its pricing and controls line up with that audience. The API is pay-as-you-go at $15 per million UTF-8 bytes with no subscription. The current S2.1-Pro model covers 83 languages and takes inline direction tags such as [warm], [near-whisper], [chuckle] and [long pause], so a companion's delivery can follow the mood of the conversation without a separate emotion parameter. It streams over WebSocket, with official Python and TypeScript SDKs.

Its other draw is a large community voice library and fast cloning, which make it easy to prototype a cast. For anything you ship, use voices you own or have cloned with consent: community voices vary in quality and rights clarity. Voice design from a text description costs $0.01 per successful request, which is cheap enough to generate many candidates for each character.

The longer-term reason to consider Fish Audio is the open-weight path. Its fish-speech, S1 and S2 models are released as open weights under a paid commercial license, and enterprises can self-host them in their own cloud or on-prem. Self-hosting is priced as an enterprise engagement (from $10,000 a month), so it isn't for a small team, but it is insurance against the kind of shutdowns that hit Replica and Play.ht users. Watch the concurrency tiers: the API starts at 5 concurrent requests and rises to 15 and 50 after $100 and $1,000 in prepaid spend.

Microsoft Azure Neural TTS is the least flashy option here and the most practical for two situations: you need speech to work offline or on-device, or your studio already runs on Azure. Embedded speech, a limited-access feature you apply for, lets supported voices run locally without a network connection, which removes network latency entirely and keeps a game playable on a plane. Few cloud-first vendors on this list offer anything comparable.

For style control Azure relies on SSML. Many neural voices support speaking styles such as cheerful, sad, angry, whispering or shouting, with adjustable intensity, plus fine control of pitch, rate, pauses and pronunciation. That is more verbose than inline tags like [laugh], but it is deterministic, which some dialogue tools prefer. The Speech SDK has C# and C++ bindings that fit Unity and Unreal projects, and real-time synthesis streams audio as it is produced.

Pricing is predictable and low at volume: $15 per million characters for neural and HD Flash voices pay-as-you-go, with commitment tiers down to $7.50 per million, and 0.5 million free characters a month. Custom voices cost more. A professional custom voice is $24 per million characters, plus training (up to $936 per training run) and endpoint hosting at $4.04 per model per hour, which adds up if every character gets its own model. The stock voices also sound more like polished assistants than actors, so Azure fits background NPCs, accessibility narration and on-device fallbacks better than hero characters.

Deepgram earns its place by owning both ends of the voice loop. Most voiced characters need to hear the player as well as speak, and Deepgram's Flux speech-to-text model was built for turn-taking, with end-of-turn events that help you decide when the character should start talking. Pairing it with Deepgram's own Aura-2 or Flux TTS keeps both legs on one API key, one bill and one set of regions, and the Voice Agent API can run STT, an LLM and TTS over a single WebSocket.

Aura-2 is priced at $0.030 per 1,000 characters on pay-as-you-go ($30 per million), and Flux TTS at $0.045. Concurrency is solid for a self-serve account, at up to 45 concurrent TTS connections on pay-as-you-go and 60 on Growth. New accounts get $200 in free credit, enough for a meaningful load test. Enterprises can self-host, which helps if you run game servers in your own data centers.

The limitation for this use case is character range. Deepgram's voices are tuned for clear, professional conversation, the kind you want from a support agent rather than an orc warlord, and it lacks the voice design and inline acting tags that ElevenLabs, Inworld and Fish Audio offer. It's a good fit for tutors, guides and companion personas with a natural, grounded voice, and for teams that value one reliable vendor for the whole voice pipeline. For a deeper look at the listening side, see our guide to speech-to-text APIs for developers.

Rime publishes some of the most aggressive latency numbers in the market. Its Mist v3 model lists 37 ms P50 and 56 ms P90 time to first audio at single concurrency, and about 70 ms when self-hosted on your own hardware. The Coda model trades some speed (96 ms P50) for more natural expression and adds word-level timestamps, which are useful for lip-sync and for cutting audio cleanly when a player interrupts. Both stream over HTTP and WebSockets.

Two features stand out for game teams. Deterministic pronunciation control renders names and alphanumerics the same way every time, which helps when your world is full of invented words. Self-hosting via Docker Compose or Kubernetes on the Enterprise plan lets you run the model next to your game servers, removing a network hop and keeping player audio inside your own infrastructure.

Pricing is simple usage billing: $0.03 per 1,000 characters for Mist v3 and $0.05 for Coda, with free starting minutes and 20 concurrent generations on the self-serve plan. The reason Rime ranks last here is fit, not quality. Its voices, case studies and tuning target contact centers in healthcare, finance and hospitality, and Mist v3 speaks only English, French, German and Spanish. If you need a calm, fast, natural voice for a guide or companion character, it is worth testing. If you need a snarling villain, look higher on this list.