[Bug]: Improvement of weam voice feature

[rutvi-weam-ai]

Overview

The voice-to-text feature fails to transcribe long audio recordings (10–15 minutes or more) reliably. This issue appears to stem from improper handling of long-duration audio — potentially due to missing logic around chunking, buffering, or transcription submission. As a result, transcriptions are often incomplete or entirely missing, which affects users relying on long-form voice inputs.

---

Affected Endpoint and Files

• Endpoint(s): (Not applicable – client-side feature issue)
• Route file: (N/A)
• Controller: (N/A)
• Middleware (if applicable): (N/A)
• Service/DAO: (N/A)
• Client Components:
  • nextjs/src/components/Chat/VoiceChat.tsx
  • nextjs/src/components/Chat/ChatClone.tsx

---

Evidence

• Voice recordings longer than ~10–15 minutes often fail to return a full transcription.
• Either:
  • No transcription is added to the chat, or
  • Only the beginning portion of the audio is transcribed, with the rest missing.
• Users report this issue consistently during long recordings.

---

Steps to Reproduce

• Open the chat interface in the app.
• Start a voice recording session.
• Speak continuously for 10–15 minutes.
• Stop the recording and wait for transcription to complete.
• Check the resulting text in the chat window

---

Expected Behavior

• Transcription is missing or incomplete.
• Only the first few minutes are transcribed, or the chat receives no text at all.
• Users do not receive feedback or error messages.

---

Proposed Remediation

• Support audio durations of 10–15+ minutes in both capture and transcription.
• If using audio chunking, ensure all chunks are received, stored, and concatenated correctly.
• Check transcription submission logic for failures with large inputs.
• Add temporary debug logs to identify where data is lost (recording, buffering, upload, transcription).

[Nisheet-Patel]

I’m working on this issue; I will update once I make progress.

[krishnamurthy38952-dotcom]

I want to contribute please assign me

[kstij1]

Hey @Nisheet-Patel , Any updates on this? let us know if you need any help.

[Nisheet-Patel]

The issue stems from the browser's native SpeechRecognition API, which automatically stops after a period of inactivity or during long transcriptions due to built-in timeouts.

My fix is to automatically restart the voice recognition service whenever it stops on its own. It will only stop permanently when the user manually clicks the stop button. This should solve the issue with long recordings.

I've successfully tested this auto-restart logic in a standalone HTML/JS file, and it works as expected for long-duration recordings.

I am going to apply these changes to the Weam codebase and am open to any suggestions or feedback on this approach.

console screenshot