`SpeechToText.transcribe()` drops preceding 30-second chunks on audio > 30s when `verbose: false`

[xgiovio]

Description

When transcribing audio waveforms longer than 30 seconds (e.g., 60s recording), calling transcribe(waveform, { verbose: false }) drops/discards all preceding 30-second windows and returns only the final 30-second window segment.
Passing { verbose: true } resolves the issue and returns the complete transcription across all 30-second windows.

Steps to reproduce

1. Record or generate a PCM 16kHz Float32Array audio waveform longer than 30 seconds (e.g., 60 seconds of speech).
2. Execute sttModel.transcribe(waveform, { language: 'en', verbose: false }).
3. Inspect result.text.

Snack or a link to a repository

No response

React Native Executorch version

latest

React Native version

0.86.2

Platforms

iOS

JavaScript runtime

None

Workflow

None

Architecture

None

Build type

None

Device

None

Device model

No response

AI model

No response

Performance logs

No response

Acknowledgements

Yes

[msluszniak]

Hi @xgiovio, thank you for submitting an issue. We will investigate it. cc: @IgorSwat

[IgorSwat]

Hey @xgiovio , could you please provide some more details, like the device you are running it on, or preferably the exact audio that caused the bug? Because I tested a bunch of >30s and >60s audios, and I was not able to replicate it ☹️

Also, could you verify that you actually use the latest version (that is v0.9.3) of the package?

[xgiovio]

there is some inconsistency, right now on an iphone 17 pro max, latest ios, latest executorch, react native 0.86.2, expo 57 latest packages i have this problem.
I'm not streaming, just recording an audio text from youtube, the audio is saved in memory and passed in full to transcribe. The output in verbose or non verbose is just the last window. If instead i segment the audio in multiple windows of 20 seconds i have for each of them the output and then i stick them myself, but is something that the model should do oh his own.
Now, i didn't check very deep, but the problem could be in the constant kMaxDecodeLength = 128 where a model can generate max 128 tokens for 29 sconds of audio. Problem i think is in ASR.cpp and is a discard trigger caused on lines 341-344 and 63-66

[IgorSwat]

@xgiovio Yeah, you are right, the mentioned piece of code is fragile, will be fixed.
Although I am still not 100% sure it responsible for this particular issue, since exceeding the 128 token limit for 30s chunk would require a really high speech pace, with high word/s ratio and no pauses. We will investigate it further.