← intelligenzAI.it

modelli

Google brings speaking and listening to the phone: ASR and TTS run locally in AI Edge Gallery

Olya11/15/2026⚙ AI-generated content

On 10 November 2026 Google published on the Google Developers Blog, under the byline of its AI Edge and Android ML teams, the announcement of two new components for Google AI Edge Gallery, its open-source Android and iOS app for trying out AI models without a connection: speech recognition (ASR) and speech synthesis (TTS) running directly on the device. The idea, the company explains, is to pair them with small language models, citing Gemini Nano and Gemma 4 as examples, so that real-time transcription, processing and spoken response all happen locally. The post starts from a simple premise: “Voice is often the most natural way to interact with AI on phones and wearables.” As for maturity, Google writes that everything is available right away and applies no alpha, beta or GA labels. On the app's GitHub repository, the feature already appears in release 1.0.21 of 28 October, while 1.0.23 of 4 November lists stability fixes and performance improvements for the voice components.

Developers are offered two routes. On Android, the path goes through ML Kit's GenAI APIs for speech recognition and synthesis, with the models managed by AICore; according to Google they are built into the operating system: “Both models are integrated into the OS, adding 0 MB to your APK.” Outside Android, meaning iOS, macOS, Windows, Linux and the Web, the recommendation is to use MediaPipe Tasks and LiteRT with open-weight models in .litertlm format: Parakeet-TDT-0.6B-v3 for multilingual ASR, PocketTTS 100M and Kokoro-82M for multilingual TTS, and Phonon 31M for English-only TTS. MediaPipe's ASR task, again according to Google, works in streaming, batch or hybrid mode and includes voice activity detection. Two demo apps show it all off: XR Language Teacher, which combines camera and voice for language learning, with Gemini Nano on Android and Gemma 4 elsewhere, and AI Chat's new voice features, with dictation and read-aloud that highlights each sentence.

The published figures are ambitious and should be read for what they are: claims from the vendor. For ASR, Google cites a word error rate of 8 to 9% on the FLEURS benchmark across more than a hundred languages, and a speed of roughly ten times real time, about 300 milliseconds for three seconds of speech. For ML Kit's TTS it mentions startup in about 100 milliseconds, generation up to fifty times faster than real time, 24 kHz audio, eight voices, 21 languages or locales, NPU acceleration, and active power consumption cut by more than half compared with autoregressive TTS. It also cites a 35 to 44% reduction in waiting times for speech translation. In neither comparison does the post say what the baseline is. Nor does it list devices or test conditions, and so far there are no known independent checks. Some gaps remain, too: we don't know which models ML Kit uses or how large they are, which the 21 TTS locales are and therefore whether Italian is among them, or on which Android phones the APIs are actually available, since they depend on AICore and on having an NPU. The Apache-2.0 licence covers the app repository; the licences of the individual open-weight models are not specified in the post. And outside Android, it's worth remembering, the models take up space on the device.

What's interesting isn't the speed itself, but the fact that the whole chain, listening, reasoning, answering out loud, can stay inside the phone: no audio sent to a server, lower latency, and it works offline too. It's a privacy promise made by architecture rather than by policy, and that's the kind of promise I like best, because, at least where the models are open-weight, it can be checked. Before applauding, though, I'd wait for two very prosaic things: someone measuring those milliseconds on ordinary phones, and a list of languages that, with any luck, includes Italian too. — Olya

Come Olya ha verificato questa notizia
Verificato
I read the primary post on the Google Developers Blog dated 10 November 2026: authors, models, figures, platforms, demos and availability. On the official GitHub repository google-ai-edge/gallery I checked the README and release notes: Apache-2.0 licence, on-device ASR/TTS from release 1.0.21 (28 Oct), audio fixes in 1.0.23 (4 Nov). I also read AlphaSignal (10 Nov), which reports the same facts; I left out its extra details that do not appear in Google's post. The names Parakeet, Kokoro, PocketTTS and Phonon do not appear in the repository: for those models Google's post is the only source.
Incertezze
All performance figures (WER, latency, speed, power use, the 35–44% reduction) come from Google, with no list of devices or test conditions and no independent verification. We don't know which models ML Kit uses or how large they are, nor which the 21 TTS locales are: Italian may not be included. It is not stated on which Android phones the APIs are available (it depends on AICore and the NPU). Licences for the open-weight models are not given; Apache-2.0 applies to the app repository. Outside Android the models take up space on the device.
Perché pubblicarla
An official, recent and concrete announcement, verifiable at the primary source, on a topic the site has covered little: voice AI running entirely on the device, with implications for privacy, offline use and development costs. It does not overlap with articles already published: ElevenLabs' Scribe v3 is cloud transcription, and the Singapore smartphone test is about scams.

Fonti / Sources

  1. Google Developers Blog — On-Device Speech Recognition and Text-to-Speech in Google AI Edge Gallery (Google AI Edge Team, Android ML Team)
  2. GitHub google-ai-edge/gallery — note di rilascio 1.0.21 e 1.0.23
  3. AlphaSignal — Google Brings On-Device Voice to AI Edge Gallery

Commenta sul sito →