Speech guide · Sources reviewed October 7, 2026

Choose a speech provider by workflow.

A managed voice agent, an audiobook and an offline kiosk need different controls. These source-linked guides cover current candidate families and the tests that decide between them.

This is a selection guide based on linked primary documentation. It does not report a new CodeSOTA benchmark run or a complete ranking of every release.

Speech evidence hub →

Build the right shortlist

  • Real-time agents: Sonic 3.6, Eleven v4 Turbo / Flash v2.5, Deepgram Aura-2 and native Gemini 3.8 Live. Compare complete turn latency, transcript fidelity and interruption behavior.
  • Audiobook narration: Eleven v4 / Multilingual v2, Azure HD, Chirp 3 HD and Gemini 3.8 TTS. Test chapter-level continuity and the exact pronunciation and SSML controls supported by your voice.
  • Local deployment: Kokoro, Piper, Fish Audio, Dia, F5-TTS and other named checkpoints. Review code and weight terms separately.

Compare evidence before prices

The previous vendor table selected each provider's maximum MOS from an unverified mixed-study catalogue. That is not a common naturalness ranking and has been removed. Use identical prompts, voice conditions, volume normalization and a blind listener protocol.

Check current plan prices, billing units, request limits, output codecs, regions and commercial voice rights against the chosen model. A price per character, per token and per minute cannot be compared without an explicit conversion workload.

Catalogue and verification status · Measured evidence status · Separate preference study.