Skip to content

TRANSCRIPTION_FIX.md

Latest commit

 

History

History
48 lines (37 loc) · 1.42 KB

TRANSCRIPTION_FIX.md

File metadata and controls

48 lines (37 loc) · 1.42 KB

PCM Format Correction

Issue

Previous fix used WAV format, but Gemini documentation states that raw 16-bit PCM at 16kHz is the optimal input format.

Correction

Reverted to raw PCM format with the correct MIME type specification.

Changes Made

Updated gemini-client.js

MIME Type (Line 76)

  • Correct format: audio/pcm;rate=16000
  • The rate=16000 parameter is crucial - it tells Gemini the sample rate

Audio Data (Line 60)

  • Sending raw PCM data (Int16Array buffer)
  • Little-endian byte order (default for JavaScript)
  • 16-bit samples
  • Mono channel
  • 16kHz sample rate

Removed WAV Methods

  • Removed createWavBuffer() method (not needed)
  • Removed writeString() helper (not needed)
  • Sending raw PCM is simpler and optimal per Gemini docs

Gemini Audio Specifications

According to official documentation:

  • Optimal format: Raw 16-bit PCM, little-endian, mono, 16kHz
  • MIME type: audio/pcm;rate=16000
  • Token usage: 32 tokens per second of audio
  • Downsampling: Gemini downsamples to 16 Kbps automatically
  • Max length: 9.5 hours per prompt

Why This Should Work Better

  1. Using the exact format Gemini recommends
  2. Proper MIME type with rate parameter
  3. No unnecessary WAV wrapper overhead
  4. 2-second chunks for better context

Testing

  1. Reload the extension
  2. Try dictating
  3. Should now get accurate transcriptions

The raw PCM format is what Gemini is optimized for!