v0.3.100

RealtimeSTT 0.3.100

New VAD callbacks on_vad_start and on_vad_stop

• triggering on VAD presence
• reverted functionality of on_vad_detect_start, on_vad_detect_stop back to: triggered when the system starts/stops detecting for VAD presence

v0.3.99

RealtimeSTT 0.3.99

1. Enhanced Logging Configuration

• Introduced a dedicated named logger realtimestt instead of using the root logger.
• Added structured logging with handlers for both console (level set by user) and file (always DEBUG).
• Logging no longer propagates to the root logger by default (logger.propagate = False).

2. Added possibility to disable Faster-Whisper VAD Filter

• Added faster_whisper_vad_filter parameter (default: True) to enable voice activity detection (VAD) from the faster_whisper library.
• Improves robustness against background noise at the cost of additional GPU resources.
• Integrated into both real-time and main transcription workflows.

3. Audio Worker Improvements

• Added improved, detailed debug logging for audio device initialization, sample rate handling, and resampling.

4. VAD Callback Adjustments

• fixes #215
• Moved on_vad_detect_start and on_vad_detect_stop callbacks to trigger directly during voice activity checks instead of state transitions.
• Ensures callbacks align more accurately with actual speech/silence events.

v0.3.98

RealtimeSTT 0.3.98

• minor fix for pypi wheel

v0.3.97

RealtimeSTT 0.3.97

• fix for #210

v0.3.95

RealtimeSTT 0.3.95

• better warmup (using audio file)
• merged #200

v0.3.94

RealtimeSTT 0.3.94

• New Parameters for stop-method of AudioToTextRecorder:

  • backdate_stop_seconds (float, default=0.0):

    • Description: Specifies the number of seconds to backdate the stop time when ending a recording.
    • Usage: When invoking stop() due to a wake word detection or a speaker diarization change event, this parameter compensates for any latency, ensuring that only relevant audio is included in the recording and transcription.

  • backdate_resume_seconds (float, default=0.0):

    • Description: Specifies the number of seconds to backdate the resume time when restarting listening after a recording has stopped.
    • Usage: Typically set to the same value as backdate_stop_seconds, this parameter allows for fine-tuning.

v0.3.93

• fix for stt-server (got broken by webservers dependency upgrade because of an api change)
• added initial_prompt_realtime to AudioToTextRecorder to be able to give different prompts to final and realtime model
• added new parameters to client/server (download root, batch sizes)

v0.3.92

• fixed dependencies (causing "ImportError: cannot import name 'BatchedInferencePipeline' from 'faster_whisper'")

v0.3.91

• upgrade to 0.3.91 since 0.3.9 had issues on PyPi

v0.3.9

RealtimeSTT v0.3.9 Release Notes

🚀 New Features

Batched Transcription

• Added support for batched transcription in both main and real-time models which improves performance and efficiency
• New parameters introduced:
  • batch_size: Controls the batch size for main transcription tasks.
  • realtime_batch_size: Configures batch size for real-time transcription.

This feature is designed to speed up processing. I can't say yet if there may be cases where batching overhead impacts performance negatively. It looked promising for me in initial tests, but I need your feedback! Please report if you get into any issues or notice even slower transcription due to batching.