0.2.8

Dependency and Compatibility Changes

• Removed Triton <3 requirement
• Tested compatibility with Python 3.14 and 3.15

Performance Improvements

• Simulstreaming backend now defaults to MLX-Whisper (if available) or Faster-Whisper (if available) encoders, paired with Whisper cross-attention and decoder using an AlignAtt policy, for increased speed. Can be disabled using --disable-fast-encoder
• Encoders are loaded once and shared in Simulstreaming, reducing vRAM usage
• Only the decoder of Whisper is loaded when using a different encoder, reducing vRAM usage

Frontend Enhancements

• Added a microphone picker
• Loads the UI as a single inline HTML file (instead of separate CSS, JS, SVGs and HTML files) for simplified deployment

Bug Fixes and Improvements

• Resolved warmup error when no connection is provided or when the language is set to auto
• Added pip timeout and retries in Dockerfile when installing Torch/TorchVision/TorchAudio
• Fixed issue where an exception is raised when language is set to 'auto' and task is set to 'translation'
• Enabled auto-detection of language for warmup if not specified

0.2.7

0.2.7: Diarization Improvements

• New default backend: Sortformer is now the default diarization backend, replacing Diart
• 6x faster processing: Reduced latency from ~2s to ~0.3s on CPU
• Significantly improved speaker detection (Constraint: Currently supports up to 4 speakers maximum)
• Shared model loading: A single Sortformer model SortformerDiarization is now shared across users and instances to reduce memory footprint. Speaker caches, frames, etc. are handled per user in SortformerDiarizationOnline
• Enhanced alignment: Improved time and token synchronization between transcription and diarization results

0.2.6

• Voice Activity Control (VAC) by Default: VAC is now enabled by default to improve transcription accuracy by filtering out non-speech segments before processing transcription & diarization. You can disable it with the --no-vac flag.

• Simulstreaming Backend Enhancements:

  • The simulstreaming backend is now the default transcription backend.
  • Improved timestamp accuracy for audio segments longer than 30 seconds.
  • Backends models are now recycled to optimize resource usage, by removing whisper hooks at the end of a transcription
  • Added the ability to preload multiple backend models using the --preloaded_model_count argument, when several users are espected

• Diarization with Silences: The diart diarization backend now correctly handles pauses and silences, improving speaker turn detection.

• Time Handling: Aligned time handling between the backend and the frontend for better synchronization.

• WebSocket Communication: Buffering is disabled during silent periods.

• Default Model: The default model is now base.

0.2.5

Build & Dependencies

• Migrated to pyproject.toml - Replaced setup.py with PEP-recommended packaging bda72b8
• Removed NumPy version constraint - No longer restricted to numpy < 2.0.0 197293e

Backend Architecture

• Refactored SimulStreaming backend separation - Improved architecture to allow multiple users to share the same backend Whisper model instance d098af3 197293e
• Enhanced performance monitoring - Lag metrics now update every 0.1 seconds and are independent of token emission frequency 2bbdc70
• Reduced hallucinations - SimulStreaming is now less likely to generate false transcriptions during silent periods 87b9ed6

Frontend Improvements

• Enhanced silence indicators - Now displays three distinct types of silences:
  • Model-detected silences ([BLANK_AUDIO])
  • Token emission gaps
  • End-of-transcription silences
    38b4ebe
• Dark theme support - Added dark mode 4e56130
• Improved UX during transcription by @davidgumberg
  • Screen no longer goes to sleep while transcribing 7f93c4b
  • Auto-scroll to latest transcription text 3b96fb8

0.2.4

Bug Fixes

• Diarization Queue Audio Overlap Fixed a bug where diarization_queue was sent the entire self.pcm_buffer on every iteration, instead of just the latest chunk. PR by @choomegan (commit)

• License Display Error Fixed dual license warning display when using simulstreaming backend. 46efbdf

Enhancements

• Improved Punctuation Splitting for Diarization Enhanced the use_punctuation_split logic to improve diarization results. Commits: 3ad3683, 5b9977c, 56114d3

• Deployment Guide Update Fixed and clarified the Deployment Guide in the README. PR by @luisla-rivas (commit)

• Architecture Update e40b5a3

• Dockerfile Improvements Updated Dockerfile to install build-essential and update the PyTorch version. - (Idea from @callumgarven) (commit)

Core Updates

• Update to latest version of SimulStreaming Fixes warmup with >30s audio files

• SimulStreaming Whisper Core Update Updated SimulStreaming whisper core from version 20230918 to 20250625. Solves tensor mismatch on some gpus due to triton version - Commits: 8e056cb, 4cfed6e

0.2.2

New:

• Replace ffmpeg-python with raw ffmpeg calls:

  • Fixes systematic crashes after 9 minutes on some machines
  • Improves reboot and restart handling
  • Allows ffmpeg to restart without crashing the server on conversion errors

• Update to latest SimulWhisper:

  • Adds compatibility with English-only models
  • Infers word-level timestamps for better diarization alignment
  • Other improvements: https://github.com/ufal/SimulStreaming/commits/main/

• Prevent buffer from growing indefinitely when no tokens are created

• Fix Hugging Face token file handling in Docker

• Remove default 8000 port in WebSocket when no port is provided

0.2.1

New SimulStreaming backend for transcription. Associated preprint: https://arxiv.org/abs/2506.17077

• Up to 5 time faster on tiny model: #134 (comment)
• Requires to install pip install whisperlivekit[simulstreaming]. Dual licensed: https://github.com/ufal/SimulStreaming?tab=readme-ov-file#-licence-and-contributions
• Use it with --backend simulstreaming
• SimulStreaming limitations for now:
  • No buffer preview is available
  • Diarization maybe be less precise
  • Punctuation can be less accurate
  • English-only model (tiny.en, base.en, medium.en) are not compatible for now

0.1.9

Faster Diarization, Smarter Speaker Splitting

• Faster diarization, with buffering logic and fixed-size audio chunks, now aligned with --min-chunk-size for improved real-time performance
• Punctuation-based speaker splitting (beta): enables more natural transitions using --punctuation-split
• Custom diarization models: use --segmentation-model and --embedding-model to specify alternate backends. See here to get a list of available models

0.1.8

Changed

• TranscriptionEngine (ex WhisperLiveKit ) can now be initialized with parameters directly via its constructor (e.g., TranscriptionEngine(backend="faster-whisper", model="small"), for greater flexibility for programmatic use in addition to command-line argument parsing.

Moved

• New module whisperlivekit.parse_args for handling command-line argument parsing.
• New module whisperlivekit.web.web_interface for serving the web interface HTML.

0.1.7 -> 0.1.8: 993a835

0.1.7

Changelog - v0.1.7

Bugs Fixes

• Fixed #127 : Transcription with VAC is functional again

Enhancements

• Backend logs now update the lag value even when no audio is detected
• Added explicit error message when ffmpeg is not found
• Frontend now indicates when no audio is detected