Topic
Text-to-speech synthesis is technology that converts written text into spoken audio, used in voice assistants, transcription services, and accessibility tools.
Fish Audio released a 35B mixture-of-experts speech model on Oct 13 that scores 93.2% on Artificial Analysis' pronunciation benchmark. Free commercial use stops at $500,000 in annual revenue, and the two-RTX-3090 setup is one Reddit user's report.
Reality
Microsoft's MAI-Transcribe-2 covers 60 languages with speaker labels and word timestamps, though streaming is not among its documented features. Live voice products get a fast model for replies and still need another way to hear the caller.
Perspective Coverage
8 publishersReality
Google shipped two text-to-speech models with the same per-line performance control and put voice creation on only one of them, so the tier decision depends on whether a project has to invent a voice at all.
Perspective Coverage
7 publishersReality
LiveKit Agents now speaks in Microsoft's MAI voices, including MAI-Voice-2-Flash, through an official plugin that covers text-to-speech only. Teams no longer write their own adapter, but every deployment still needs an Azure Speech key and region.
Reality
ElevenLabs released Eleven v4 Turbo for voice agents, reporting medians of about 100 ms inference latency and 150 ms to first speech. Moving an existing agent takes more than a model ID swap, since older voice clones need retraining and SSML break tags no longer work.
Perspective Coverage
3 publishersReality
One developer's lip-sync estimator measured its first test vowel against a running average dominated by that vowel, across three TTS voices. Rotating which vowel leads, from a reset state, is the control a stateful audio model needs before its score means anything.
Reality
Loquent's 14,000 vox-bench runs found Claude Haiku's latency tail at 2.85x its median while ElevenLabs tightened to 1.92x. Alert thresholds tuned to June's data would now be watching the wrong stage of the pipeline.
Reality
Gemini 3.8 Flash TTS and Flash-Lite TTS arrive with matching APIs, more than 2,000 stock voices and a path from a 30-second sample to a replica voice. Securing the speaker's permission is the customer's job.
Reality
Breeze TTS 2 scores Elo 1215 and lists at $34 per 1M characters, about a third of Eleven v3's rate. Only the inference code is Apache 2.0, and commercial use of the weights needs a written agreement from BreezeBlue.
Reality
OpenBMB's 2B-parameter model is Apache-2.0, speaks 30 languages, and serves through vLLM's OpenAI-compatible /v1/audio/speech. The real-time factor quoted for it was measured on a different backend than the one the project recommends for production.
Reality
In a dev.to walkthrough, FFmpeg's silencedetect finds the gaps, a word budget prices each one at 2.5 words a second, and any narration that renders longer than its gap is thrown out and generated again shorter.
Reality
Apple's Kids Category rule and COPPA's definition of a child's voice both point a preschool voice product at local processing, and Whisper's error on children's speech falls by nearly a factor of three from tiny to large-v3.
Reality
The audio platform says AI now produces 99% of its new content and made production about 80 times cheaper, and the number worth watching is 12-month revenue retention, which it puts at 76%.
Reality
A dev.to writeup put caption-driven diffusion TTS against a Style-Bert-VITS2 baseline on the same GPU slice and found a fixed overhead that alone exceeds the baseline's entire generation time, so the model moved offline.
Reality