Background Transcription

Background transcription processes audio chunks in real-time while recording continues, providing immediate feedback and eliminating wait time at the end of long recordings.

What is Background Transcription?

When VAD chunking is enabled, Voicepad transcribes each chunk in the background as soon as it's extracted:

graph TB
    subgraph "Recording Thread"
        A[Record Audio] --> B[Audio Queue]
    end

    subgraph "Chunk Worker Thread"
        B --> C[VAD Analysis]
        C --> D{Chunk Ready?}
        D -->|No| B
        D -->|Yes| E[Extract Chunk]
        E --> F[Transcribe Chunk]
        F --> G[Append to Markdown]
        G --> H[Accumulate Audio]
    end

    A --> I[Continue Recording...]
    H --> J[Next Chunk...]

Key principle: Recording and transcription happen simultaneously, not sequentially.

How It Works

1. Chunk Detection

When VAD detects a natural speech boundary:

Chunk audio is extracted from the buffer
Chunk is added to the transcription queue
Recording continues uninterrupted

2. Background Transcription Thread

A dedicated worker thread:

✅ Monitors queue for new chunks
✅ Transcribes each chunk using Whisper model
✅ Appends result to markdown file (thread-safe)
✅ Accumulates audio data in memory
✅ Continues until recording stops

3. Model Caching

The first chunk load the Whisper model into memory. Subsequent chunks reuse the cached model:

# First chunk: Load model (~2-5 seconds)
Chunk 1: [Load Model] → [Transcribe 60s audio in 3s]

# Subsequent chunks: Reuse cached model (~0 seconds overhead)
Chunk 2: [Transcribe 60s audio in 3s]
Chunk 3: [Transcribe 60s audio in 3s]

Result: Only the first chunk has loading overhead!

4. Thread-Safe Writing

Multiple chunks may be transcribing simultaneously (if transcription is fast). A lock ensures markdown file updates are sequential and never corrupted:

with markdown_lock:
    append_to_file(chunk_transcription)

Real-Time Markdown Updates

Live Progress

Open the markdown file while recording to watch transcription appear:

# Transcription: meeting_20260218_033000.wav

**Status:** Recording in progress...

---

## Chunk 1 (0:00 - 1:15)

Welcome everyone to today's meeting. We'll be discussing...

## Chunk 2 (1:15 - 2:45)

[This appears while you're still recording chunk 3]

First on the agenda is the project timeline...

Live Viewing

Use a markdown viewer with auto-refresh to see transcription update live. VS Code's markdown preview auto-updates when the file changes.

Completion Status

When recording stops, the status updates:

**Status:** Recording complete ✓

---

**Recording Complete**
- Total chunks: 5
- Total duration: 8:25

Audio File Output

Memory Accumulation

Chunk audio data is kept in memory (not saved individually):

accumulated_chunks = [
    chunk_1_audio,  # 60 seconds
    chunk_2_audio,  # 58 seconds
    chunk_3_audio,  # 62 seconds
]

Single Merged File

When recording stops, all chunks merge into one WAV file:

merged_audio = concatenate(accumulated_chunks)
save_to_file("recording_20260218_033000.wav")

Result: One continuous audio file, no chunk files on disk.

Performance

Transcription Speed

Speed depends on device and model:

GPU (CUDA):

Tiny model: 20-40x real-time (60s → 1.5-3s)
Medium model: 10-20x real-time (60s → 3-6s)
Large model: 5-10x real-time (60s → 6-12s)

CPU:

Tiny model: 3-8x real-time (60s → 7.5-20s)
Medium model: 1-3x real-time (60s → 20-60s)
Large model: 0.5-1x real-time (60s → 60-120s)

CPU with Large Models

On CPU, large models may transcribe slower than real-time. This means chunks can pile up faster than they're processed, causing delays.

Optimal Settings

For smooth background transcription:

# GPU users (fast transcription)
vad_min_chunk_duration: 60.0
transcription_model: medium
transcription_device: cuda

# CPU users (slower transcription)
vad_min_chunk_duration: 120.0    # Larger chunks, less frequent
transcription_model: tiny         # Faster model
transcription_device: cpu

Error Handling

Transcription Failures

If a chunk fails to transcribe:

✅ Error logged with details
✅ Worker continues processing next chunk
✅ Recording is not interrupted
✅ Failed chunk is still saved in audio file

The markdown will show:

## Chunk 3 (2:30 - 3:45)

[Transcription failed for this chunk]

Worker Resilience

The chunk worker is designed to never crash:

try:
    transcribe_chunk()
except Exception as e:
    log_error(e)
    continue  # Process next chunk

Even if transcription fails completely, you still get the audio file!

Monitoring Progress

Log Output

Watch the terminal for real-time progress:

[INFO] Chunk 1 ready (62.3s), transcribing...
[INFO] Chunk 1 transcribed successfully
[INFO] Chunk 2 ready (59.8s), transcribing...
[INFO] Chunk 2 transcribed successfully

Chunk Worker Status

On stop, see summary:

[INFO] Chunk worker stopped. Processed 5 chunks, accumulated 5 total chunks
[INFO] Saving 5 accumulated chunks...
[INFO] Recording saved to: data/recordings/meeting_20260218_033000.wav
[INFO] Total duration: 8:25.6 seconds

Comparison: With vs Without Background Transcription

Traditional (No VAD)

Timeline: [Record 10 min] → [Stop] → [Transcribe 2 min] → [Done]

Total time: 12 minutes
Wait time: 2 minutes

Output timing:
- Audio: Available immediately
- Transcription: Available after 2 minutes