8 Best Text-to-Speech APIs of 2026

Cartesia builds Sonic, a text-to-speech model based on state space models, and markets it on speed: the vendor documents sub-90 ms model latency and streaming that starts playback before the full reply is generated. Sonic 3.6, released in August 2026, speaks 44 languages, reads confirmation codes and heteronyms without preprocessing, and adds Hinglish support. Developers stream over a WebSocket that accepts partial transcripts through contexts and continuations, which suits LLM output that arrives token by token. Official Python and JavaScript SDKs plus LiveKit, Pipecat and Twilio guides shorten setup. Pricing runs on credits at 1 credit per character: Free gives 20K a month, Pro costs $5 for 100K, Startup $49 for 1.25M and Scale $299 for 8M, with overage from $65 down to $38 per million. Instant voice cloning needs only 10 seconds of audio. Language coverage is narrower than ElevenLabs.

ElevenLabs sells its text-to-speech API on pay-as-you-go pricing per 1,000 characters, with subscription plans from $6 a month for teams that prefer bundles. The model menu is the broadest in this list. Eleven v4 is the expressive flagship with audio tags for fine control, v4 Turbo keeps most of that expressiveness at about 100 ms median latency, and Flash v2.5 is the budget real-time option for 32 languages with a 40,000 character limit per request. Both v4 models cover 90+ languages. Fast models cost $0.04 per 1,000 characters and the expressive ones $0.08, so a voice agent that runs on v4 Turbo pays $40 per million characters. The pricing page also lists a startup grant of 33 million characters over 12 months. The downside is cost at volume: Inworld, Deepgram and the cloud providers charge less per character, so heavy users should model their monthly totals before committing.

Deepgram started in speech-to-text and now sells three text-to-speech models on the same per-character pricing. Flux TTS is its conversation model: it reads the whole exchange so tone and pacing carry across turns, handles interruptions in the model, and starts audio in as low as 80 ms at 24kHz. It costs $0.045 per 1,000 characters. Aura-2 at $0.030 is the volume workhorse, and Aura-1 at $0.015 is the cheapest line. Flux TTS ships with built-in text normalization for English plus 10 languages, including Spanish, Hindi and Mandarin, and supports 40+ more if you normalize text yourself. Streaming runs over WebSocket and batch over REST. Deepgram also lists integrations with Twilio, LiveKit, Daily, Pipecat and jambonz, offers BAAs for eligible customers, and supports self-hosted deployment. The $200 sign-up credit is generous for testing.

Inworld sells two realtime text-to-speech models. Realtime TTS-2 Flash is built for latency, with first audio as low as 50 ms, and costs $15 per million characters on the On-Demand plan. Realtime TTS-2 is the expressive model, taking delivery directions in plain words either for the whole request or in brackets mid-line, at $25 per million characters. Inworld says one TTS-2 voice speaks about 500 languages and dialects, and both models render 50 non-verbal sounds. Both stream over WebSocket and HTTP, and an OpenAI-compatible endpoint lets you point the OpenAI SDK at Inworld with few code changes. Monthly plans from $25 to $1,500 buy credits and cut rates, down to $7 per million characters for TTS-2 Flash on Growth. Cloned and designed voices cost the same as stock voices. Compliance add-ons such as HIPAA sit on the top self-serve tier, and no on-prem option is listed.

Rime focuses on voices for customer conversations rather than narration. It offers two text-to-speech models: Mist v3 for the lowest latency at $0.030 per minute on pay as you go, and Coda v2 for the most natural delivery at $0.050 per minute. Rime estimates about 1,000 characters per minute, so Mist v3 works out near $30 per million characters. The Free plan includes $50 of usage every month, and the $100 Scale plan cuts every rate by 20%. Rime lists sub-100 ms time to first byte and 600+ voices across 50+ languages. Its docs cover streaming over HTTP and WebSockets, LiveKit and Pipecat quickstarts, and phoneme control for Coda with custom lexicons so names and brand terms come out the same every time. Enterprises can license Coda or Mist v3 engines to run on their own NVIDIA hosts. The per-minute unit makes cost comparison less direct.

Google Cloud Text-to-Speech is a mature API with a long list of voice families. Standard and WaveNet voices cost $4 per million characters after a free 4 million characters a month, Neural2 costs $16, Chirp 3: HD costs $30 after a free million, and Studio voices $160. Chirp 3: HD is the family Google aims at agents, with conversational voices and low-latency streaming. Gemini-TTS adds prompt-based control over style and is billed per token, with preview models on a separate schedule. Google lists 380+ voices across 75+ languages and variants, SSML support and long audio synthesis of up to 1 million bytes of input. Instant custom voice costs $60 per million characters. Developers call it over REST or gRPC with Google's client libraries. The trade-off is setup: you need a Google Cloud project, billing and IAM, and Google does not publish a single latency figure.

Azure AI Speech bundles text-to-speech with transcription, translation and avatars in one Azure resource. Standard neural voices cover 100+ languages and locales and cost $15 per million characters, with a free tier of 0.5 million characters a month. Professional custom voice synthesis costs $24 per million characters, or $48 on neural HD, plus training at $52 per compute hour up to $936 per training and $4.04 per model per hour for hosting. Custom voice is a limited-access feature, so you need approval. SSML support is deep: custom lexicons, speaking styles and language switching on multilingual voices. Developers use the Speech SDK, REST API or Speech CLI, and Speech Studio offers a no-code tool for tuning output. Visemes help sync lips for avatars, and Azure sells its own avatar at $0.50 per minute for real-time. Setup is heavier than with the startups, and Microsoft publishes guidance on lowering latency rather than one figure.

Amazon Polly is the long-running AWS speech service and still one of the cheapest. Standard voices cost $4 per million characters, Neural voices $16, Generative voices $30 and Long-Form voices $100. New accounts get a free tier for the first 12 months, including 100 thousand Generative characters a month. Polly streams audio back as soon as you send text and supports common SSML tags plus Amazon extras such as a Newscaster speaking style on selected voices. Speech Marks return timing data for lip sync and karaoke-style highlighting, and custom lexicons fix how names are read. Polly works through every AWS SDK, from Python to C++, and integrates with Amazon Connect, Genesys Cloud CX and the Amazon Chime SDK. For an exclusive voice, AWS runs a Brand Voice engagement with your team. The voice catalog is less expressive than ElevenLabs or Cartesia, and AWS publishes no latency figure.