# Best text to speech APIs for voice agents

Which text to speech APIs stream over a WebSocket, keep one voice across languages, return word timing, and send 8 kHz phone audio, per each vendor's docs.

Published 2026-11-17. Paxa Labs.

A voice agent needs a text to speech API that speaks while the language model is still writing and stops when the caller interrupts. Paxa TTS Flash v2, Deepgram Aura, Cartesia, Fish Audio, Botnoi Voice, Murf, and ElevenLabs each document a streaming connection for that. The languages your callers speak and the audio format your phone system expects decide which of them stay on the list.

Every vendor fact below links that vendor's own documentation, as [our comparison pages](https://paxalabs.com/compare/text-to-speech-apis) verified it. No vendor is scored or ranked, and each vendor's published rate is on that comparison.

| Service | Live connection, as documented | Timing and phone formats | Languages its pages name |
| --- | --- | --- | --- |
| Paxa TTS Flash v2 | A WebSocket that takes text fragments as they arrive, with a cancel frame for interruptions | Word, sentence, or utterance timing. G.711 μ-law and A-law at 8 kHz | English, Thai, and Mandarin in every voice |
| Deepgram Aura | A WebSocket that queues Speak messages and returns audio after each Flush | linear16, mulaw, and alaw on the socket | Seven languages on Aura-2, with no Thai or Mandarin |
| Cartesia | A websocket endpoint for realtime synthesis | An add_timestamps parameter | 44 on Sonic 3.6, Thai among them |
| Fish Audio | A WebSocket that buffers text until it has enough context, with a flush event | The pages cited state none | S2 Pro names Thai among its added languages |
| Botnoi Voice | A realtime websocket that returns audio in base64 chunks | No timing field in its API reference | More than 20, with Thai, English, and Chinese |
| Murf | Falcon 2, its streaming model for conversational AI and voice agents | ALAW and ULAW, with 8000 Hz among its sample rates | 35, with a Thai section in its voice library |
| ElevenLabs | HTTP and websocket access | ulaw_8000 and alaw_8000 output formats | Thai on Eleven v4, v4 Turbo, v3, and v3 Conversational |

## Which text to speech APIs support WebSocket streaming for voice agents?

[Deepgram's streaming guide](https://developers.deepgram.com/docs/streaming-text-to-speech) describes a WebSocket that queues text sent in Speak messages and returns audio after a Flush message. [Cartesia's API reference](https://docs.cartesia.ai/api-reference/tts/tts) has a websocket endpoint for realtime synthesis. [Fish Audio's streaming guide](https://docs.fish.audio/features/realtime-streaming) describes a WebSocket that buffers text until it has enough context, with a flush event that forces generation.

[Botnoi Voice's API reference](https://voice.botnoi.ai/developer/api) documents a realtime websocket that returns audio in base64 chunks. [Murf's API overview](https://murf.ai/api/docs/introduction/overview) names Falcon 2 as its streaming model for conversational AI and real-time voice agents. [ElevenLabs' API reference](https://elevenlabs.io/docs/api-reference/introduction) offers HTTP and websocket access.

The Paxa [live WebSocket](https://paxalabs.com/docs/live-streaming) stays open for a whole conversation. The agent sends text fragments as the language model writes them, and audio frames come back as soon as each part is ready. A flush frame speaks whatever is buffered at the end of a turn. A cancel frame stops the reply when the caller talks over it and keeps the connection open for the next turn.

A reply that is already complete needs no socket. A streamed HTTP request returns audio in chunks while the rest is still being made, and [the latency guide](https://paxalabs.com/docs/latency) explains when each connection reaches first audio sooner. [What an AI voice agent is](https://paxalabs.com/guides/what-is-an-ai-voice-agent) covers the listening and thinking steps around the voice.

## Which text to speech APIs let one voice speak many languages?

Every Paxa TTS Flash v2 voice speaks English, Thai, and Mandarin, including an English product name inside a Thai sentence. The agent keeps one voice when a caller switches language halfway through a call. [The voices page](https://paxalabs.com/voices) plays all 50.

Language lists differ by vendor and by model. [Deepgram's Aura voices](https://developers.deepgram.com/docs/tts-models) cover seven languages, and neither Thai nor Mandarin is one of them. [Cartesia](https://docs.cartesia.ai/build-with-cartesia/tts-models/latest) names Thai among the 44 languages of Sonic 3.6. [ElevenLabs' models page](https://elevenlabs.io/docs/overview/models) lists Thai for Eleven v4, Eleven v4 Turbo, Eleven v3, and Eleven v3 Conversational. [MiniMax's API](https://platform.minimax.io/docs/api-reference/speech-t2a-http) sets Thai, English, Chinese, and Cantonese as separate language values, and [Gemini-TTS](https://docs.cloud.google.com/text-to-speech/docs/gemini-tts) lists Thai as generally available with Mandarin in Preview.

A language list says the service accepts the text. Whether one voice sounds like the same speaker in each language is something to hear on your own script. [The best Chinese text to speech guide](https://paxalabs.com/guides/best-chinese-text-to-speech) and [the best Thai text to speech guide](https://paxalabs.com/guides/best-thai-text-to-speech) go through each language by job.

## Which APIs return word timing for captions and interruptions?

An agent that shows live captions, or needs to know which words the caller heard before interrupting, needs timing with the audio. Paxa TTS Flash v2 returns [word, sentence, or utterance timing](https://paxalabs.com/docs/timestamps) on HTTP requests and on the live connection, where the spans arrive as events beside the audio frames. Thai text has no spaces, and the word spans split it into words.

[Cartesia's API](https://docs.cartesia.ai/api-reference/tts/tts) documents an `add_timestamps` parameter. [Microsoft's Speech SDK](https://learn.microsoft.com/en-us/azure/ai-services/speech-service/how-to-speech-synthesis) raises a WordBoundary event at each spoken word, with its time offset in the audio. Botnoi Voice's API reference documents no timing field. [The word-level timestamps guide](https://paxalabs.com/guides/word-level-timestamps-api) explains what the spans are used for.

Code written for OpenAI's speech endpoint can call [Paxa's compatible endpoint](https://paxalabs.com/openai-compatible-api) by changing the base URL and the key. Word timing is served on Paxa's own speech endpoint and live connection, under the same key.

## Which APIs send audio a phone line can play?

Telephony carries audio as 8 kHz μ-law or A-law. Paxa TTS Flash v2 returns both directly, over HTTP and the live connection alike. The audio goes into a telephony stack with no resampling step. It also returns MP3, Opus, WAV, and raw PCM at 8 to 48 kHz for apps and browsers.

[Deepgram's output settings](https://developers.deepgram.com/docs/tts-media-output-settings) list linear16, mulaw, and alaw for its streaming WebSocket. [ElevenLabs' create speech reference](https://elevenlabs.io/docs/api-reference/text-to-speech/convert) lists `ulaw_8000` and `alaw_8000` and notes that μ-law is common for Twilio audio inputs. [Murf](https://murf.ai/api/docs/text-to-speech/overview) lists ALAW and ULAW outputs with 8000 Hz among its sample rates, and [its voice library](https://murf.ai/api/docs/voices-styles/voice-library) has a Thai section.

## How do you retry a speech request without paying for the same audio twice?

A timeout leaves the client unsure whether the request ran. On Paxa's HTTP speech endpoints, send an [Idempotency-Key header](https://paxalabs.com/docs/idempotency) with each request. A retry with the same key and the identical text is served from the first attempt's record and charged once.

Derive the key from something the agent already has, such as the call id and the turn number. For [rate limits](https://paxalabs.com/docs/rate-limits), read the retry delay the API returns and wait that long before sending again.

## What does one spoken turn cost?

Paxa TTS Flash v2 bills by the character at 10 credits per 1,000 characters, which is $10 per million. A reply of 200 characters costs 2 credits, about $0.002. With timing on, the rate is 1.25 times that, and the same reply costs 2.5 credits, about $0.0025.

A new account gets 100 free credits with no card, for personal use and evaluation. Vendors price by characters, by minutes, or by plan, and [the comparison](https://paxalabs.com/compare/text-to-speech-apis) lists each published rate with its source.

## When does another service fit better?

Paxa Labs offers no voice cloning, and Paxa TTS Flash v2 speaks three languages. An agent that must speak Spanish, German, French, Dutch, Italian, or Japanese is covered by Deepgram's Aura-2. An agent that needs a clone of a specific speaker should look at the vendors that document cloning, which [the Thai buyer's guide](https://paxalabs.com/guides/best-thai-text-to-speech) lists. A product already built on one vendor's socket protocol also carries a real cost to move.

## What should you read or try next?

- See the live connection on [the realtime text to speech page](https://paxalabs.com/realtime-tts) and the product on [the text to speech page](https://paxalabs.com/text-to-speech).
- Hear a reply in your agent's voice in [the Speech tool](https://paxalabs.com/app/speech).
- Connect an agent framework through [the LiveKit page](https://paxalabs.com/livekit-text-to-speech) or [the OpenAI-compatible API](https://paxalabs.com/openai-compatible-api).
- Read [the live streaming guide](https://paxalabs.com/docs/live-streaming), [the latency guide](https://paxalabs.com/docs/latency), and [the idempotency guide](https://paxalabs.com/docs/idempotency).
- Compare published rates on [the text to speech API comparison](https://paxalabs.com/compare/text-to-speech-apis).
- Read how teams build them on [the voice agents page](https://paxalabs.com/use-cases/voice-agents).
- For the listening side, read [realtime and recorded transcription](https://paxalabs.com/guides/realtime-vs-recorded-transcription) and [the multilingual speech to text guide](https://paxalabs.com/guides/best-speech-to-text-api-for-multilingual-audio).

## Common questions

### What is the best text to speech API for a voice agent?

The one that speaks your callers' languages over a connection your stack supports. Paxa TTS Flash v2 fits agents that speak English, Thai, or Mandarin in one voice, over a WebSocket or a phone line. Deepgram Aura, Cartesia, Fish Audio, Botnoi Voice, Murf, and ElevenLabs document streaming connections too. Test two or three on your own script.

### What is real-time speech-to-speech voice AI?

It is a system that hears a person and answers aloud within the same conversation. Most are built from three parts: speech to text writes down what the caller said, a language model writes the reply, and text to speech speaks it. The text to speech part decides how soon the caller hears the first word.

### Can real-time text to speech voice a game or companion character?

Yes. The same live connection that serves a phone agent serves a character, and Paxa TTS Flash v2 reads written directions such as [whispering] or [laughs] inside the line. [The AI companion page](https://paxalabs.com/ai-companion) covers that use.

### What is the cheapest text to speech API?

It depends on the unit each vendor bills and on how much you send. Paxa TTS Flash v2 costs $10 per million characters on every plan. [The comparison](https://paxalabs.com/compare/text-to-speech-apis) converts each vendor's published rate to the same unit, with its source.

### How do you test a Thai voice agent?

Write the replies your agent really gives, with an amount in baht, a date, a phone number, and an English brand name. Listen on a handset, where callers hear it. Thai text has no full stop. Check that the agent ends each turn with a flush. [The sentence breaks guide](https://paxalabs.com/guides/thai-sentence-breaks-for-streaming-speech) shows where to cut a Thai reply.
