8 Text-to-Speech APIs for Voice Agents Compared

The best text-to-speech API for a voice agent is the one that streams audio fast on your network, sounds natural on phone audio, and pronounces your names, numbers and terms correctly.

Vendor latency claims are a starting point, not a result; measure time to first audio in your own stack.

This comparison covers eight APIs using official documentation read on 3 November 2026.

How We Evaluated

We did not test these APIs. Each was reviewed against the same criteria using official docs:

• Streaming: WebSocket or chunked streaming for real-time playback

• Latency positioning: what the vendor documents, labeled as a vendor claim

• Voice options: stock voices, cloning or custom voices

• Telephony fit: output formats such as 8 kHz mu-law for phone calls

• Deployment: cloud regions, enterprise or self-hosted options

The Eight Options at a Glance

1. ElevenLabs

ElevenLabs offers multiple models, including Flash models aimed at low latency, plus WebSocket streaming and voice cloning.

• Strength: broad voice library and multilingual support

• Watch for: cost at high call volume; compare plan limits

2. Cartesia Sonic

Cartesia positions Sonic for real-time conversational use with streaming over WebSockets.

• Strength: designed for interactive voice agents

• Watch for: confirm language coverage for your markets

3. Deepgram Aura

Deepgram Aura provides streaming TTS aimed at voice agents, alongside Deepgram's speech recognition.

• Strength: one vendor for STT and TTS

• Watch for: voice variety compared with voice-library vendors

4. OpenAI TTS

OpenAI's text-to-speech guide documents a speech endpoint with streaming and promptable voice style on newer models.

• Strength: simple API for teams already on OpenAI

• Watch for: fixed set of built-in voices; no custom cloning in the standard API

5. Azure AI Speech

Azure AI Speech offers neural voices, SSML control and custom neural voice under limited access.

• Strength: fine control over pronunciation and prosody via SSML

• Watch for: custom voice requires an application and approval

6. Google Cloud Text-to-Speech

Google Cloud TTS provides several voice families, including Chirp HD voices, and streaming synthesis.

• Strength: many languages and Google Cloud integration

• Watch for: feature support varies by voice type

7. Amazon Polly

Amazon Polly offers standard, neural, long-form and generative engines, with SSML and lexicons.

• Strength: fits AWS and Amazon Connect deployments

• Watch for: voice availability differs by engine and region

8. Rime

Rime focuses on voices for conversational and phone use, with streaming APIs.

• Strength: voices tuned for call-center style speech

• Watch for: confirm enterprise terms and language coverage

Our own product. LuMay's Voxentis voice agent works with several TTS providers, including ElevenLabs and Deepgram, chosen per deployment. We are not ranked here. See the Voxentis voice agent.

What to Measure Yourself

• Time to first audio: from text sent to first audio byte, measured in your region

• Full turn latency: speech end to agent audio start, including STT and LLM

• Pronunciation: names, addresses, product codes, dates and currency

• Phone audio quality: listen at 8 kHz, not only on studio headphones

• Interruption handling: how quickly playback stops when the caller speaks

• Cost per call minute: at your expected characters per call

Related reading: speech-to-text APIs for voice agents and how to test voice agents for latency and handoffs.

Selection Checklist

• Shortlist three APIs by language and deployment needs

• Build a 50-line test script with your real names, numbers and terms

• Measure time to first audio from your hosting region

• Run blind listening tests with staff on phone audio

• Check voice cloning consent rules and terms of use

• Compare cost at expected monthly volume

Limitations

Latency and pricing change often. We relied on vendor docs and did not benchmark. Not verified: any specific millisecond figure.

Next Step

If you need a production voice agent with the right TTS for your callers, the Voxentis voice agent team can help select and tune a provider.

• Cartesia documentation

• Deepgram text-to-speech

• OpenAI text-to-speech guide

• Azure AI Speech text to speech

• Google Cloud Text-to-Speech

• Amazon Polly

• Rime documentation