ML Kit

The ML Kit GenAI text-to-speech library is currently in alpha, available on some Pixel and Samsung devices, offering 8 voices and support for 21 language locales, and runs entirely on-device after the first download.

Gemini Nano 4 is built on the open-source model Gemma 4 and is invoked directly on supported flagship Android devices through the ML Kit API; the model is delivered by the Android system, so apps do not need to bundle the model.

The ML Kit GenAI speech recognition API is in alpha; the basic mode supports Android devices at API level 31 and above, while the advanced mode is powered by Gemini Nano and provides better quality and language coverage on Pixel 10 and above devices.

ML Kit mainly solves common problems in the vision and natural language domains, such as text recognition, face detection and translation.

The GenAI API uses Gemini Nano to enable capabilities such as text summarization and image description, with inference performed entirely on-device.

MediaPipe spans Android, iOS, Web and Python platforms, supports model customization and is open source; ML Kit supports only Android.

ML Kit imposes an upper limit on Gemma 4's context length, so developers must first confirm the budget to fully use 128K.

In an end-to-end scam detection case, generating a response took less than 120 milliseconds on most devices.

ML Kit adds speech translation capability, supporting transcription in more than 20 languages.

More than 200,000 apps mention AI in their app store descriptions, with downloads expected to approach 10 billion over the same period.

Constrained decoding is a technique that forces the model's output to conform to a specific grammar or format during inference, masking to 0 the probability of tokens that do not conform to a predefined grammar (such as JSON Schema) at each token generation step.

When choosing tools, the usual first question is whether the task requires a custom model: general tasks prioritize ready-made APIs like ML Kit, and when fine-tuning on private data or domain-specific accuracy is needed, then take the custom model path of Core ML or TensorFlow Lite.

ML Kit wraps a set of out-of-the-box APIs on top of TensorFlow Lite, such as text recognition, barcode scanning, face detection, and language identification, suitable for teams that do not want to train models from scratch.

In a tech talk, Google engineers Amy and Alice introduced building audio-first AI experiences for smart glasses using Gemini Nano 4 and ML Kit, and provided an open-source sample, XR Language Teacher.