Voice AI blog.
Models, research and ideas.

Explore the latest in voice AI and text-to-speech, with model releases, research and technical insights from the Neuphonic team.

10/11/2026

We sent 432 requests from London to test NeuTTS-Max, our cloud text-to-speech model. The median wait for speech was 83 ms, including network time and any silence before the first word. Here’s how we measured it.

How we measure TTS latency from request to first audible speech

10/11/2026

How we measure TTS latency from request to first audible speech

10/11/2026

Our cloud TTS API benchmark, from request to first audible speech.

How we measure TTS latency

10/11/2026

Introducing NeuTTS-Max, our most capable text-to-speech model, available through our cloud API for real-time voice agents. Explore 44 languages, IPA pronunciation control and word-level timestamps, plus the benchmark behind its 83 ms median time to speech.

NeuTTS-Max: Low-Latency TTS API, 83 ms Time to Speech

10/11/2026

NeuTTS-Max: Low-Latency TTS API, 83 ms Time to Speech

10/11/2026

Our cloud TTS API model with 44 languages, pronunciation control and word timestamps.

Introducing NeuTTS-Max

08/10/2026

NeuDecide turns speech directly into structured tool calls, without a transcription step. We introduce our small decision models, explain how they select tools and fill in arguments, and show what this enables for robotics, wearables and smart devices.

NeuDecide: An audio-first tiny decision model

08/10/2026

NeuDecide: An audio-first tiny decision model

08/10/2026

End-to-end tool selection and argument prediction from speech.

NeuDecide: An audio-first tiny decision model

11/09/2026

Our neural audio codec is now available through Hugging Face Transformers. We explain how we built the integration, the design decisions behind it, and how to get started with the AutoModel API.

NeuCodec in Transformers 5.17

11/09/2026

Our neural audio codec is now available through Hugging Face Transformers. We explain how we built the integration, the design decisions behind it, and how to get started with the AutoModel API.

NeuCodec in Transformers 5.17

11/09/2026

Integration details and usage through the AutoModel API.

NeuCodec in Transformers 5.17

30/07/2026

NeuTTS-2E brings emotional control to a compact, 125M-parameter text-to-speech model. With seven emotion modes and four preset voices, it generates expressive speech privately and offline - even when the words contain no emotional cues.

NeuTTS-2E: Emotional Control in On-Device Speech

30/07/2026

Meet NeuTTS-2E, our compact on-device model for private, controllable emotional speech across seven emotion modes - even when the text has no emotional cues.