Download on iOS & Android

Features

• Transcription – Turn voice messages into text instantly
• Summarization – Skip the "uhh" & "ahh" and get to the key points
• Translation – Break language barriers with a single tap
• Auto-detection – Get notified when new messages arrive

Tech Stack

• Flutter (Hive, GoRouter, Riverpod, Pigeon for cleaner method channels)
• Groq (Whisper)
• DeepL for translation
• Fastlane for publishing
• Firebase (Analytics, Crashlytics, Storage for user feedback)

How It Works (Expand section below for more details)

• The app checks for new voice message files when opened (manual trigger available).
• Transcriptions are displayed in a list with a summary option.
• Users can enable notification listening for automatic detection.
• If enabled (or if the app is backgrounded), a notification with the transcription is sent.
• On iOS, users must share the audio file manually since files are not programmatically accessible.

Detailed Version

Finding the Audio Files

WhatsApp stores two types of audio files:

1. Voice Notes (standard WhatsApp voice messages)
2. Audio Files (shared audio not recorded in WhatsApp)

Trim Talk focuses on voice notes, which are stored under: /storage/emulated/0/WhatsApp/Media/WhatsApp Voice Notes/ Each week’s messages are stored in folders formatted as YEAR-WEEKNUMBER. However, WhatsApp does not follow ISO week numbering, requiring a workaround:

• If the expected week's folder is missing, the app checks the previous week instead.

Reading the Files

Since these files are in another app's dedicated folder, access varies by Android version:

• Android 12 and below: Requires READ_EXTERNAL_STORAGE permission.
• Android 13+: Three options, but only Storage Access Framework (SAF) works.
  • Media Store (not possible due to .nomedia file preventing indexing).
  • MANAGE_EXTERNAL_STORAGE (restricted to file manager apps).
  • SAF (requires user selection of the folder and accessing files via content resolver).

SAF is not well-supported in Flutter. Existing packages failed, so I wrote custom method channel calls to access files.

Transcribing the Files

Initially, I aimed for on-device transcription, testing multiple solutions:

• Android Speech-to-Text API (incompatible with audio files).
• Whisper (various implementations: TensorFlow, Mediapipe, method channels, etc.)
  • None provided an optimal balance of performance and accuracy due to mobile hardware limitations.

Cloud-based APIs were tested:

• Deepgram, Google, AssemblyAI, OpenAI Whisper
  • Worked but were slow, inaccurate, or expensive.
• Groq
  • Uses an LPU™ Inference Engine, accelerating open-source models like whisper-large-v3.
  • Fast, accurate, and cost-efficient.

Automating the Process

To avoid requiring manual checks:

• Used workmanager for background tasks (every 15 min), transcribing files and displaying notifications.
• However, method channels do not work in background tasks due to separate isolates.
• Tried multiple alternatives without success (likely possible natively, but not in Flutter).
• Best workaround: Notification Listener Service to trigger processing (not ideal due to reliability and permissions required).

---

Key Takeaways

• Flutter is powerful but has limitations with native/platform-specific features.
• Android's developer experience is not always fun.
• Groq is an excellent transcription solution.
• Deep understanding of isolates, method channels, and the Flutter engine.
• Experience in building, publishing, and maintaining Flutter packages and apps.

Code & Commands

Useful Commands

• Generate Hive: dart run build_runner build -d
• Generate Pigeon API: dart run pigeon --input pigeon_api.dart
• Check the scripts folder (fix Pods, publish, etc.)

Debug Mode

• Notifications always appear.
• WorkManager debug notifications are visible.
• Transcriber returns dummy data.
• read file permission is always true (android)

Links

Download on iOS and Android or share the Universal Link.

Support

I built this project on my free time, if you'd like to support it, consider contributing here. Thank you! :)