[Feature request] Add word-level timing metadata and short-utterance guidance for GenAI Speech Recognition

[SoturiSV]

What's your feature request? Please describe.

We are evaluating com.google.mlkit:genai-speech-recognition:1.0.0-alpha1 for a production Android app that performs live, structured voice-command transcription from short utterances in noisy environments.

We would like ML Kit GenAI Speech Recognition to expose richer recognition metadata. The current alpha API appears to expose only text: String on PartialTextResponse and FinalTextResponse. For production use, we need word-level metadata such as:

• token/word text
• start/end offsets
• confidence scores, if available
• alignment/timing data for partial and final results

Without timing/alignment data, apps must synthesize word timing, which is too error-prone for workflows that need to group words into phrases and correlate recognition with VAD/audio events.

We also need documented guidance on the intended audio/session pattern for short utterances, including:

• recommended minimum clip duration
• recommended lead/tail padding
• silence handling
• endpointing behavior
• whether short session-based recognition is supported
• whether long continuous streams are the intended usage pattern

Mobile environment

Android

ML Kit API/version: com.google.mlkit:genai-speech-recognition:1.0.0-alpha1

Additional context

Related Issue Tracker feature request: https://issuetracker.google.com/issues/540021698

In our testing with real live-use audio replayed through AudioSource.fromPfd():

• A single long speech-only stream continued receiving audio for the full stream but recognized only about 48 words from a 1,234-word expected transcript.
• Separate VAD-derived clips as individual recognizer sessions improved coverage slightly to about 119 words, but most clips returned ERROR_TYPE_NO_SPEECH_DETECTED.

This appears related to existing reports about AudioSource.fromPfd() sensitivity to silence/endpointing, but this request is broader: richer response metadata plus guidance on the intended usage pattern for short utterance recognition.