[BUG] Parakeet v3 transcription fails with ORT error for sessions longer than ~5 minutes

Bug Description Parakeet v3 transcription fails with ORT error for sessions longer than ~5 minutes. Short sessions (under 5 min) transcribe successfully, but sessions of 10+ minutes consistently fail. Current Behavior Recording stops successfully and audio file is saved completely. When transcription begins for sessions ~10 minutes or longer, it fails with: transcription failed: transcription engine failed: parakeet transcription failed: ORT error For sessions under 5 minutes, transcription completes successfully in 1-15 seconds. Expected Behavior Parakeet v3 should transcribe sessions of any duration (10-20 minutes) without error, consistent with how Parakeet v2 previously worked before v0.2.1. Steps to Reproduce

Install Meetily v0.2.1 on Windows Go to Settings → Transcription → select Parakeet (app downloads parakeet-tdt-0.6b-v3-int8 automatically) Record a session under 5 minutes → transcription succeeds Record a session of 10+ minutes → transcription fails with ORT error Verify the saved audio file is complete and healthy — the bug is in transcription only, not recording

Environment

OS: Windows Browser: N/A (desktop app) Version: v0.2.1 Node Version: Unknown npm/pnpm Version: Unknown

Screenshots/Videos N/A Error Messages transcription failed: transcription engine failed: parakeet transcription failed: ORT error Additional Context

The recorded audio file is confirmed complete and healthy for all session lengths — the issue is purely in the transcription engine Meetily v0.2.1 only downloads parakeet-tdt-0.6b-v3-int8 — v2 is no longer offered by the app Previously (before v0.2.1), Parakeet processed audio in small real-time chunks throughout the session. In v0.2.1, v3 processes the entire session as one chunk after recording stops — this whole-chunk approach appears to hit a memory or buffer limit in ONNX Runtime beyond ~5 minutes of audio Verified model files are intact: encoder-model.int8.onnx (~652MB), decoder_joint-model.int8.onnx (~17MB), nemo128.onnx (~139KB), vocab.txt CPU and memory usage shows no spikes during recording or transcription Reinstalling the app does not fix the issue as model files persist in C:\Users<user>\AppData\Roaming\com.meetily.ai\models\parakeet
Switching to Whisper model in Meetily results in transcription getting stuck at "waiting to generate transcript"

Possible Solution

Reintroduce chunk-based processing for Parakeet v3 as was done in previous versions, rather than whole-session single-chunk processing — this would avoid hitting ONNX Runtime memory/buffer limits for longer sessions Alternatively, re-offer Parakeet v2 as a fallback option in the model selection settings Investigate ONNX Runtime session memory limits when processing audio buffers beyond a certain duration threshold on Windows

Checklist

I have searched for similar issues I have provided all required information I have included screenshots/videos if applicable I have included error messages if applicable