Recent Releases of stt
stt - Coqui STT 1.4.0
General
This is the 1.4.0 release for Coqui STT, the deep learning toolkit for speech-to-text. In accordance with semantic versioning, this version is backwards compatible with previous 1.x versions. The compatibility guarantees of our semantic versioning cover the deployment APIs: the C API and all the official language bindings: Python, Node.JS/ElectronJS and Java/Android. You can get started with Coqui STT 1.4.0 by following the steps in our documentation.
Compatible pre-trained models are available in the Coqui Model Zoo.
We also include example audio files:
which can be used to test the engine, and checkpoint files for the English model (which are identical to the 1.0.0 checkpoint and provided here for convenience purposes):
coqui-stt-1.4.0-checkpoint.tar.gz
which are under the Apache 2.0 license and can be used as the basis for further fine-tuning. Finally this release also includes a source code tarball:
Under the MPL-2.0 license. Note that this tarball is for archival purposes only since GitHub does not include submodules in the automatic tarballs. For usage and development with the source code, clone the repository using Git, following our documentation.
Notable changes
Added experimental WebAssembly support
With the new WASM package you can deploy Coqui STT directly in the browser.
Added ARMv7 and AArch64 Python wheels for Python 3.7 and 3.9
Migrated .NET bindings to .NET Framework 4.8
Rewritten audio processing logic in iOS demo app
Documentation
Documentation is available on stt.readthedocs.io.
Contact/Getting Help
- GitHub Discussions - best place to ask questions, get support, and discuss anything related to 🐸STT with other users.
- Gitter - You can also join our Gitter chat.
- Issues - If you have discussed a problem and identified a bug in 🐸STT, or if you have a feature request, please open an issue in our repo. Please make sure you search for an already existing issue beforehand!
Contributors to 1.4.0 release
- Alessio Placitelli
- Anton Yaroshenko
- ChamathKB
- Ciaran O'Reilly
- Daniel Souza
- Danny Waser
- David Roundy
- Davidian1024
- Edresson Casanova
- Josh Meyer
- Mariano Gonzalez
- NanoNabla
- Reuben Morais
- Yanlong Wang
We’d also like to thank all the members of our Gitter chat room who have been helping to shape this release!
- C++
Published by github-actions[bot] over 3 years ago
stt - Coqui STT 1.3.0
General
This is the 1.3.0 release for Coqui STT, the deep learning toolkit for speech-to-text. In accordance with semantic versioning, this version is backwards compatible with previous 1.x versions. The compatibility guarantees of our semantic versioning cover the deployment APIs: the C API and all the official language bindings: Python, Node.JS/ElectronJS and Java/Android. You can get started today with Coqui STT 1.3.0 by following the steps in our documentation.
Compatible pre-trained models are available in the Coqui Model Zoo.
We also include example audio files:
which can be used to test the engine, and checkpoint files for the English model (which are identical to the 1.0.0 checkpoint and provided here for convenience purposes):
coqui-stt-1.3.0-checkpoint.tar.gz
which are under the Apache 2.0 license and can be used as the basis for further fine-tuning. Finally this release also includes a source code tarball:
Under the MPL-2.0 license. Note that this tarball is for archival purposes only since GitHub does not include submodules in the automatic tarballs. For usage and development with the source code, clone the repository using Git, following our documentation.
Notable changes
Added new experimental APIs for loading Coqui STT models from memory buffers
This allows loading models without writing them to disk first, which can be useful for dynamic model loading as well as for handling packaging in mobile platforms
Added ElectronJS 16 support
Rewritten audio processing logic in iOS demo app
Added pre-built binaries for iOS/Swift bindings in CI
With these two changes we're hoping to get more feedback from iOS developers on our Swift bindings and pre-built STT frameworks - how can we best package and distribute the bindings so that it feels native to Swift/iOS developers? If you have any feedback, join our Gitter room!
Extended the Multilingual LibriSpeech importer to support all languages in the dataset
Supported languages: English, German, Dutch, French, Spanish, Italian, Portuguese, Polish
Exposed full metadata information for decoded samples when using the coquisttctcdecoder Python package
This allows access to the entire information returned by the decoder in training code, meaning experimenting with new model architectures doesn't require adapting the C++ inference library to test your changes.
Added initial support for Apple Silicon in our pre-built binaries
C/C++ pre-built libraries are universal, language bindings will be updated soon
Added support for FLAC files in training
Documentation
Documentation is available on stt.readthedocs.io.
Contact/Getting Help
- GitHub Discussions - best place to ask questions, get support, and discuss anything related to 🐸STT with other users.
- Gitter - You can also join our Gitter chat.
- Issues - If you have discussed a problem and identified a bug in 🐸STT, or if you have a feature request, please open an issue in our repo. Please make sure you search for an already existing issue beforehand!
Contributors to 1.3.0 release
- Alessio Placitelli
- Danny Waser
- Erik Ziegler
- Han Xiao
- Reuben Morais
We’d also like to thank all the members of our Gitter chat room who have been helping to shape this release!
- C++
Published by github-actions[bot] almost 4 years ago
stt - Coqui STT 1.2.0
General
This is the 1.2.0 release for Coqui STT, the deep learning toolkit for speech-to-text. In accordance with semantic versioning, this version is backwards compatible with previous 1.x versions. The compatibility guarantees of our semantic versioning cover the deployment APIs: the C API and all the official language bindings: Python, Node.JS/ElectronJS and Java/Android. You can get started today with Coqui STT 1.2.0 by following the steps in our documentation.
Compatible pre-trained models are available in the Coqui Model Zoo.
We also include example audio files:
which can be used to test the engine, and checkpoint files for the English model (which are identical to the 1.0.0 checkpoint and provided here for convenience purposes):
coqui-stt-1.2.0-checkpoint.tar.gz
which are under the Apache 2.0 license and can be used as the basis for further fine-tuning. Finally this release also includes a source code tarball:
Under the MPL-2.0 license. Note that this tarball is for archival purposes only since GitHub does not include submodules in the automatic tarballs. For usage and development with the source code, clone the repository using Git, following our documentation.
Notable changes
- Added Python 3.10 support
- Added new inference APIs which process any pending data before returning transcription results
- Added an importer for using data from Common Voice's new personal data downloader, and a Jupyter notebook which creates a custom STT model using your data
- Improved and extend evaluatetflite script (now evaluateexport module) with Opus support
- Added support for Ogg/Vorbis encoded audio files as training inputs
- Added an importer for the Att-HACK dataset
- Model dimensions are now loaded automatically from a checkpoint if present
- Checkpoint loader will now handle CuDNN checkpoints transparently, without an explicit flag
- When starting a training run, a batch size check will be performed automatically to help diagnose memory issues early
- Added support for using WebDataset for training datasets
- Updated to TensorFlow Lite 2.8, including new XNNPACK optimizations for quantized models
Documentation
Documentation is available on stt.readthedocs.io.
Contact/Getting Help
- GitHub Discussions - best place to ask questions, get support, and discuss anything related to 🐸STT with other users.
- Gitter - You can also join our Gitter chat.
- Issues - If you have discussed a problem and identified a bug in 🐸STT, or if you have a feature request, please open an issue in our repo. Please make sure you search for an already existing issue beforehand!
Contributors to 1.2.0 release
- Alexandre Lissy
- Aya AlJafari
- Danny Waser
- Jeremiah Rose
- Jonathan Washington
- Josh Meyer
- Reuben Morais
- Vincent Fretin
We’d also like to thank all the members of our Gitter chat room who have been helping to shape this release!
- C++
Published by github-actions[bot] about 4 years ago
stt - Coqui STT 1.1.0
General
This is the 1.1.0 release for Coqui STT, the deep learning toolkit for speech-to-text. In accordance with semantic versioning, this version is not completely backwards compatible with previous versions. The compatibility guarantees of our semantic versioning cover the deployment APIs: the C API and all the official language bindings: Python, Node.JS/ElectronJS and Java/Android. You can get started today with Coqui STT 1.1.0 by following the steps in our documentation.
Compatible pre-trained models are available in the Coqui Model Zoo.
We also include example audio files:
which can be used to test the engine, and checkpoint files for the English model (which are identical to the 1.0.0 checkpoint and provided here for convenience purposes):
coqui-stt-1.1.0-checkpoint.tar.gz
which are under the Apache 2.0 license and can be used as the basis for further fine-tuning. Finally this release also includes a source code tarball:
Under the MPL-2.0 license. Note that this tarball is for archival purposes only since GitHub does not include submodules in the automatic tarballs. For usage and development with the source code, clone the repository using Git, following our documentation.
Notable changes
- Package missing dependencies with Android AAR packages
- Fix evaluate_tflite.py script to use new Coqpit-based config handling
- Use export beam width by default in evaluation reports
- Integrate lexicon-constrained and lexicon-free Flashlight decoders for CTC and ASG acoustic models in decoder package
- Update supported NodeJS versions to current supported releases: 12, 14, and 16
- Update supported ElectronJS versions to current supported releases: 12, 13, 14 and 15
- Improved and packaged VAD transcription module in the training package (coquistttraining.transcribe)
Documentation
Documentation is available on stt.readthedocs.io.
Contact/Getting Help
- GitHub Discussions - best place to ask questions, get support, and discuss anything related to 🐸STT with other users.
- Gitter - You can also join our Gitter chat.
- Issues - If you have discussed a problem and identified a bug in 🐸STT, or if you have a feature request, please open an issue in our repo. Please make sure you search for an already existing issue beforehand!
Contributors to 1.1.0 release
- Alexandre Lissy
- Josh Meyer
- Julian Darley
- Leon Kiefer
- Reuben Morais
- Vojtěch Drábek
We’d also like to thank all the members of our Gitter chat room who have been helping to shape this release!
- C++
Published by github-actions[bot] about 4 years ago
stt - Coqui STT 1.0.0
General
This is the 1.0.0 release for Coqui STT, the deep learning toolkit for speech-to-text. In accordance with semantic versioning, this version is not completely backwards compatible with previous versions. The compatibility guarantees of our semantic versioning cover the inference APIs: the C API and all the official language bindings: Python, Node.JS/ElectronJS and Android. You can get started today with Coqui STT 1.0.0 by following the steps in our documentation.
This release includes pre-trained English models, available in the Coqui Model Zoo:
- Coqui English STT v1.0.0-huge-vocab
- Coqui English STT v1.0.0-yesno
- Coqui English STT v1.0.0-large-vocab
- Coqui English STT v1.0.0-digits
all under the Apache 2.0 license.
The acoustic models were trained on American English data with synthetic noise augmentation. The model achieves a 4.5% word error rate on the LibriSpeech clean test corpus and 13.6% word error rate on the LibriSpeech other test corpus with the largest release language model.
Note that the model currently performs best in low-noise environments with clear recordings. This does not mean the model cannot be used outside of these conditions, but that accuracy may be lower. Some users may need to further fine tune the model to meet their intended use-case.
We also include example audio files:
which can be used to test the engine, and checkpoint files for the English model:
coqui-stt-1.0.0-checkpoint.tar.gz
which are under the Apache 2.0 license and can be used as the basis for further fine-tuning. Finally this release also includes a source code tarball:
Under the MPL-2.0 license. Note that this tarball is for archival purposes only since GitHub does not include submodules in the automatic tarballs. For usage and development with the source code, clone the repository using Git, following our documentation.
Notable changes
- Removed support for protocol buffer input in native client and consolidated all packages under a single "STT" name accepting TFLite inputs
- Added programmatic interface to training code and example Jupyter Notebooks, including how to train with Common Voice data
- Added transparent handling of mixed sample rates and stereo audio in training inputs
- Moved CI setup to GitHub Actions, making code contributions easier to test
- Added configuration management via Coqpit, providing a more flexible config interface that's compatible with Coqui TTS
- Handle Opus audio files transparently in training inputs
- Added support for automatic dataset subset splitting
- Added support for automatic alphabet generation and loading
- Started publishing the training code CI for a faster notebook setup
- Refactor training code into self-contained modules and deprecate train.py as universal entry point for training
Training Regimen + Hyperparameters for fine-tuning
The hyperparameters used to train the model are useful for fine tuning. Thus, we document them here along with the training regimen, hardware used (a server with 8 NVIDIA A100 GPUs each with 40GB of VRAM), along with the full training hyperparameters. The full training configuration in JSON format is available here.
The datasets used were: - Common Voice 7.0 (with custom train/dev/test splits) - Multilingual LibriSpeech (English, Opus) - LibriSpeech
The optimal lm_alpha and lm_beta values with respect to the Common Voice 7.0 (custom Coqui splits) and a large vocabulary language model:
- lm_alpha: 0.5891777425167632
- lm_beta: 0.6619145283338659
Documentation
Documentation is available on stt.readthedocs.io.
Contact/Getting Help
- GitHub Discussions - best place to ask questions, get support, and discuss anything related to 🐸STT with other users.
- Gitter - You can also join our Gitter chat.
- Issues - If you have discussed a problem and identified a bug in 🐸STT, or if you have a feature request, please open an issue in our repo. Please make sure you search for an already existing issue beforehand!
Contributors to 1.0.0 release
- Alexandre Lissy
- Anon-Artist
- Anton Yaroshenko
- Catalin Voss
- dag7dev
- Dustin Zubke
- Eren Gölge
- Erik Ziegler
- Francis Tyers
- Ideefixze
- Ilnar Salimzianov
- imrahul3610
- Jeremiah Rose
- Josh Meyer
- Kathy Reid
- Kelly Davis
- Kenneth Heafield
- NanoNabla
- Neil Stoker
- Reuben Morais
- zaptrem
We’d also like to thank all the members of our Gitter chat room who have been helping to shape this release!
- C++
Published by github-actions[bot] over 4 years ago
stt - Coqui STT 0.10.0-alpha.29
Test automatic release notes.
- C++
Published by github-actions[bot] over 4 years ago
stt - STT v0.10.0-alpha.7
Alpha release, for development purposes only.
- C++
Published by reuben over 4 years ago
stt - Coqui STT 0.9.3
General
This is an initial release for 🐸STT, backwards compatible with mozilla/DeepSpeech 0.9.3. The model files below are identical with to the 0.9.3 release of mozilla/DeepSpeech, and released under the MPL 2.0 license accordingly. These models are provided as a compatibility aid so that examples in our documentation can work with the existing release links.
This release includes the source code:
Under the MPL-2.0 license. And the acoustic models:
coqui-stt-0.9.3-models.pbmm coqui-stt-0.9.3-models.tflite
Experimental Mandarin Chinese acoustic models trained on an internal corpus composed of 2000h of read speech:
coqui-stt-0.9.3-models-zh-CN.pbmm coqui-stt-0.9.3-models-zh-CN.tflite
all under the MPL-2.0 license.
The model files with the ".pbmm" extension are memory mapped and thus memory efficient and fast to load. The model files with the ".tflite" extension are converted to use TensorFlow Lite, has post-training quantization enabled, and are more suitable for resource constrained environments.
The acoustic models were trained on American English with synthetic noise augmentation and the .pbmm model achieves an 7.06% word error rate on the LibriSpeech clean test corpus.
Note that the model currently performs best in low-noise environments with clear recordings and has a bias towards US male accents. This does not mean the model cannot be used outside of these conditions, but that accuracy may be lower. Some users may need to train the model further to meet their intended use-case.
In addition we release the scorer:
which takes the place of the language model and trie in older releases and which is also under the MPL-2.0 license.
There is also a corresponding scorer for the Mandarin Chinese model:
coqui-stt-0.9.3-models-zh-CN.scorer
We also include example audio files:
which can be used to test the engine, and checkpoint files for both the English and Mandarin models:
coqui-stt-0.9.3-checkpoint.tar.gz coqui-stt-0.9.3-checkpoint-zh-CN.tar.gz
which are under the MPL-2.0 license and can be used as the basis for further fine-tuning.
Training Regimen + Hyperparameters for fine-tuning
The hyperparameters used to train the model are useful for fine tuning. Thus, we document them here along with the training regimen, hardware used (a server with 8 Quadro RTX 6000 GPUs each with 24GB of VRAM), and our use of cuDNN RNN.
In contrast to some previous releases, training for this release occurred as a fine tuning of the previous 0.8.2 checkpoint, with data augmentation options enabled. The following hyperparameters were used for the fine tuning. See the 0.8.2 release notes for the hyperparameters used for the base model.
train_filesFisher, LibriSpeech, Switchboard, Common Voice English, and approximately 1700 hours of transcribed WAMU (NPR) radio shows explicitly licensed to use as training corpora.dev_filesLibriSpeech clean dev corpus.test_filesLibriSpeech clean test corpustrain_batch_size128dev_batch_size128test_batch_size128n_hidden2048learning_rate0.0001dropout_rate0.40epochs200augmentpitch[pitch=1~0.1]augmenttempo[factor=1~0.1]augmentoverlay[p=0.9,source=${noise},layers=1,snr=12~4](where ${noise} is a dataset of Freesound.org background noise recordings)augmentoverlay[p=0.1,source=${voices},layers=10~2,snr=12~4](where ${voices} is a dataset of audiobook snippets extracted from Librivox)augmentresample[p=0.2,rate=12000~4000]augmentcodec[p=0.2,bitrate=32000~16000]augmentreverb[p=0.2,decay=0.7~0.15,delay=10~8]augmentvolume[p=0.2,dbfs=-10~10]cache_for_epochs10
The weights with the best validation loss were selected at the end of 200 epochs using --noearly_stop.
The optimal lm_alpha and lm_beta values with respect to the LibriSpeech clean dev corpus remain unchanged from the previous release:
lm_alpha0.931289039105002lm_beta1.1834137581510284
For the Mandarin Chinese model, the following values are recommended:
lm_alpha0.6940122363709647lm_beta4.777924224113021
Documentation
Documentation is available on stt.readthedocs.io.
Contact/Getting Help
- GitHub Discussions - best place to ask questions, get support, and discuss anything related to 🐸STT with other users.
- Gitter - You can also join our Gitter chat.
- Issues - If you have discussed a problem and identified a bug in 🐸STT, or if you have a feature request, please open an issue in our repo. Please make sure you search for an already existing issue beforehand!
Contributors to 0.9.3 release
Everyone who helped us get this far! Thank you for your continued collaboration!
- C++
Published by reuben almost 5 years ago