TEVR ASR Tool

• state-of-the-art performance
  • 3.64% WER on Common Voice German
  • rank #1 on paperswithcode.com
• no GPU needed
• 100% offline
• 100% private
• 100% free
• MIT license
• Linux x86_64
• command-line tool
• easy to understand
  • only 284 lines of C++ code
  • AI model on HuggingFace

High Transcription Quality

In August 2022, we ranked #1 on "Speech Recognition on Common Voice German (using extra training data)" with a 3.64% word error rate. Accordingly, the performance of this tool is considered to be the best of what's currently possible in German speech recognition:

How does this work?

L175-L185 load the WAV file. L189-L229 execute the acoustic AI model. L260-L275 convert the predicted token logits into string snippets. L73-L162 implement the Beam search re-scoring based on a KenLM language model.

If you're curious how the acoustic AI model works and why I designed it that way, here's the paper: https://arxiv.org/abs/2206.12693 and here's a pre-trained HuggingFace transformers model: https://huggingface.co/fxtentacle/wav2vec2-xls-r-1b-tevr

Install the Debian/Ubuntu package

Download tevr_asr_tool-1.0.0-Linux-x86_64.deb from GitHub and extract the multipart ZIP:

wget "https://github.com/DeutscheKI/tevr-asr-tool/releases/download/v1.0.0/tevr_asr_tool-1.0.0-Linux-x86_64.zip.001"
wget "https://github.com/DeutscheKI/tevr-asr-tool/releases/download/v1.0.0/tevr_asr_tool-1.0.0-Linux-x86_64.zip.002"
wget "https://github.com/DeutscheKI/tevr-asr-tool/releases/download/v1.0.0/tevr_asr_tool-1.0.0-Linux-x86_64.zip.003"
wget "https://github.com/DeutscheKI/tevr-asr-tool/releases/download/v1.0.0/tevr_asr_tool-1.0.0-Linux-x86_64.zip.004"
wget "https://github.com/DeutscheKI/tevr-asr-tool/releases/download/v1.0.0/tevr_asr_tool-1.0.0-Linux-x86_64.zip.005"
cat tevr_asr_tool-1.0.0-Linux-x86_64.zip.00* > tevr_asr_tool-1.0.0-Linux-x86_64.zip
unzip tevr_asr_tool-1.0.0-Linux-x86_64.zip

Install it:

sudo dpkg -i tevr_asr_tool-1.0.0-Linux-x86_64.deb

Install from Source Code

Download submodules:

git submodule update --init

CMake configure and build:

cmake -DCMAKE_BUILD_TYPE=MinSizeRel -DCPACK_CMAKE_GENERATOR=Ninja -S . -B build
cmake --build build --target tevr_asr_tool -j 16

Create debian package:

(cd build && cpack -G DEB)

Install it:

sudo dpkg -i build/tevr_asr_tool-1.0.0-Linux-x86_64.deb

Usage

tevr_asr_tool --target_file=test_audio.wav 2>log.txt

should display the correct transcription mückenstiche sollte man nicht aufkratzen. And log.txt will contain the diagnostics and progress that was logged to stderr during execution.

GPU Acceleration for Developers

I plan to release a Vulkan & OpenGL-accelerated real-time low-latency transcription software for developers soon. It'll run 100% private + 100% offline just like this tool, but instead of processing a WAV file on CPU it'll stream the real-time GPU transcription of your microphone input through a WebRTC-capable REST API so that you can easily integrate it with your own voice-controlled projects. For example, that'll enable hackable voice typing together with pynput.keyboard.

If you want to get notified when it launches, please enter your email at https://madmimi.com/signups/f0da3b13840d40ce9e061cafea6280d5/join

Commercial Customization

This tool itself is free to use also for commercial use. And of course it comes with no warranty of any kind.

But if you have an idea for a commercial use-case for a customized version of this tool or for similar technology - ideally something that helps small and medium-sized businesses in northern Germany become more competitive - then please contact me at [email protected]

Research Citation

If you use this for research, please cite:

@misc{https://doi.org/10.48550/arxiv.2206.12693,
  doi = {10.48550/ARXIV.2206.12693},
  url = {https://arxiv.org/abs/2206.12693},
  author = {Krabbenhöft, Hajo Nils and Barth, Erhardt},  
  keywords = {Computation and Language (cs.CL), Sound (cs.SD), Audio and Speech Processing (eess.AS), FOS: Computer and information sciences, FOS: Computer and information sciences, FOS: Electrical engineering, electronic engineering, information engineering, FOS: Electrical engineering, electronic engineering, information engineering, F.2.1; I.2.6; I.2.7},  
  title = {TEVR: Improving Speech Recognition by Token Entropy Variance Reduction},  
  publisher = {arXiv},  
  year = {2022}, 
  copyright = {Creative Commons Attribution 4.0 International}
}

Replace the AI Model

The German AI model and my training scripts can be found on HuggingFace: https://huggingface.co/fxtentacle/wav2vec2-xls-r-1b-tevr

The model has undergone XLS-R cross-language pre-training. You can directly fine-tune it with a different language dataset - for example CommonVoice English - and then re-export the files in the tevr-asr-data folder.

Alternatively, you can donate roughly 2 weeks of A100 GPU credits to me and I'll train a suitable recognition model and upload it to HuggingFace.