v0.5.3

This PR introduces an engine system to swap Whisper engines between faster-whisper and TensorRT-LLM.

API

• MIT License!
• Added the ability to swap the Whipser "engine" from the default faster-whisper to TensorRT-LLM, which is much faster. #285
• Added support for distil models like distil-large-v2 and distil-large-v3. These work with the TensorRT-LLM engine.
• Added a batch_size parameter for the endpoints. It doesn't do anything yet, but the TensorRT-LLM engine supports batch processing of files, and the idea is add this feature along with dynamic batch.
• Overall tighter control over dependencies, and various dependency updates.

Diarization

• Started work implementing Nvidia NeMo's new long-form diarization class. Currently it still consumes too much memory.

Documentation

• Add examples for offline models for various backends #288 #289

Thanks to contributors @aleksandr-smechov and for the work from the NeMo team, and the WhisperS2T project for the initial code for the TensorRT-LLM backend, and by extension, TensorRT-LLM's Whisper example.

v0.5.2

This PR introduces a lot of things to allow remote execution and single-service deployment.

API

• Added the possibility to make RemoteExecution or LocalExecution for transcription and diarization services #258
• Implemented single-service deployment with the new only_transcription and only_diarization asr types #261
• Added new endpoints to manage remote execution servers #263
• Allow the user to auto-switch between local and remote execution for all services #266

Diarization

• Adjusted VAD speech padding and updated the diarization logic #271

Bug and Fixes

• Fixed a newly introduced dual_channel bug #268
• Added a fix to avoid empty utterances #273

Documentation

• Added new documentation to the project via mkdocs-material and GitHub pages #269

Thanks to contributors @aleksandr-smechov @chainyo

v0.5.1

API

• Added the offset_start and offset_end parameters to the API
• Add Live transcription #241 #247

Transcription

• Updated dual_channel to multi_channel with auto-detection feature #239 #244 #248
• Let faster-whisper handle the model path for the custom model hosted on HF #251

Bugs and Fixes

• Fixed the python 3.8 compat #235
• Fixed the hms format for timestamps #237
• Fixed the extra_languages config #235

v0.5.0

This release is a significant change from poetry to hatch with many improvements to CI, tests, local development, and dependencies handling.

• Huge project updates #227
• Updated tests, dependencies, config defaults #229

API

• Added a warmup for inference #201
• Added repetion_penalty parameter #207
• Added num_speakers parameter #195
• Improved the time_and_tell function #213
• Updated the API schemas #188
• Added transcription parameters for control #213

Transcription

• Added bfloat16 to compute types #209

Diarization

• Added empty audio catch during diarization #223 #225
• Reimplemented the entire diarization module to skip NeMo module installation #186 #202

CI

• Added concurrency on CI tests #191

Contributors:
@aleksandr-smechov @chainyo

v0.4.0

This release includes a lot of improvements and a new License starting with the v0.4.0 of wordcab-transcribe (inspired by the HFOIL).

The new License WTLv0.1

The new License prevents anyone from using this project after v0.4.0 (included) to sell a self-hosted version of this software without any agreements from Wordcab.

But you can still use the project for research, personal use, or even as a backend tool for your projects.

API

• Fixed CortexResponse for Svix size limit #101
• Made alignment non-critical if the process fails #105
• Added multi-GPU support for transcription, alignment, and diarization #114
• Added the audio_duration (in seconds) in the API response #127
• Added a catch for invalid or empty audio file #128
• Added a log about the number of detected and used GPUs at launch #138
• Updated pydantic to v2 #157
• Added an audio file global download queue #168
• Added the new WTL v0.1 License #177 #183 #184

Transcription

• Added the vocab feature #124
• Added an internal_vad parameter that helps with empty utterances #142 #173
• Added a new fallback for empty segments during transcription #149
• Added the float32 compute type for the transcription model #157

Diarization

• Decomposed the diarization process into sub-modules and optimized diarization inference #180

Alignment

• Added new cs, in, sl and th alignment models #164

Post-processing

• Improved the post-processing strategy #136 #157
• Fix word_timestamps parameter for dual_channel #152

Instructions

• Improvement of the contributions instructions #131

Deploy

• Update error payload for Svix in cortex endpoint #118
• Docker image updated to cuda:11.7.1 #133
• Update Svix payload in cortex endpoint #144
• Add a configuration file using Nginx for custom deploy #146

Need improvements / Not fully working

• Added the possibility to use extra transcription models for specific languages #110

Contributors:
@chainyo @aleksandr-smechov @jissagn

v0.3.1

TL;DR: Transcription is now on steroids. 2x faster than the actual faster-whisper implementation.

API

• Add time_and_tell decorator on specific functions to time individual processes on debug=True #77
• Add a LoggingMiddleware on debug=True #77
• Add a fallback for dual_channel if the audio file is not stereo #87

Transcription

• Add quality metrics for the batch process and fallback if the quality is under defined thresholds #89
• Implement word_timestamps for the batch process #91

Post-processing

• Fix timestamps format during the post-processing step #86

@chainyo

v0.3.0

Documentation

• Improve .env readability for an easier API configuration #52
• Add README instructions for profiling container #72

API

• Add authentication when the API is not in debug mode #56
• Fix the audio file endpoint inputs #59
• All submitted files are converted into .wav 16kHz for consistency #60
• Reworked and more coherent Request/Response models for the API endpoints #60
• Streamline the post-process functions (with or without alignment/diarization) #63
• Simplify timestamps conversion in outputs #63
• Fix blocking non-async functions #67
• Huge API rework for handling concurrent requests better #71
• Fix Exception/Error returns through the API -> raised errors should be more transparent for user #72
• VAD use now onnx and faster-whisper implementation #72

AI models

• Add alignment (from whisperX) as a new possible step #51
• Fix alignment for fr, de, es, and it models #59
• Add dual_channel transcription process for stereo audio file #60
• Add the choice to use diarization or not #63
• Implement Batch request process for transcription #72

Deploy

• Docker is aligned with the local setup now #55
• Improve Dockerfile and commands to use cache for models #55

Contributors:
@aleksandr-smechov @chainyo

v0.2.0

Changes from #31 @chainyo

• Replace diarization with NVIDIA NeMo asr toolkit
• Update config.py and add validators for necessary config settings
• Update the Docker image with the latest from NVIDIA
• Fix dependencies and versions
• Fix the Python version to 3.9 locally and on Docker
• New available timestamps format: ms. Now user can choose between hms, s (default) and ms.
• Remove unused num_speakers parameter.

v0.1.0

First official release:

• Open-source
• Fast
• Easy to deploy
• Batch requests
• Cost-effective
• Easy-to-use