Coqui's Fifth Mondays Newsletter
- 👩💻Work at Coqui
- 👩🍳Welcome
- 📔Jupyter Notebooks
- 🚀 Translators without Borders Deploying Coqui’s STT
- 🎥Coqui, Stars of Stage and (the Small) Screen
- Out the Door, TTS v0.1.3
- 🔬 Monthly TTS Papers
👩💻Work at Coqui
By Kelly Davis
Yeah, you heard that right; we’re hiring!
An open source remote-friendly Berlin based startup founded by the creators of Mozilla’s text-to-speech (TTS) and speech-to-text (STT) engines (over 550K downloads and 22K GitHub stars), with the backing of investors from around the globe (London, San Francisco, and Berlin), and we’re hiring!
What’s not to love?
We’re hiring across-the-board for a number of roles; so, there’s something for everyone:
- Head of Product
- 3 x Senior Full Stack Engineers
- 2 x Senior STT Deep Learning Engineers
- 2 x Senior TTS Deep Learning Engineers
- 2 x Senior, Developer Community Managers
The full list of open positions is available on our jobs page.
We’d love to hear from you; so, if any roles pique your interest, reach out to [email protected]. 🐸!
👩🍳Welcome
By Kelly Davis
This month we’ve cooked up a few tasty nuggets for you!
- To make training STT models even easier, we’ve brewed up two Jupyter notebooks to help with the process, a Colab notebook to train from scratch and a second Colab notebook to fine-tune with.
- Translators without Borders baked you a Bengali speech-to-text model that they generously released in the Model Zoo.
- Also, stars of stage and (the small) screen, we’ve stewed up a little video on our joint work with Harvard and Google.
- TTS is heating up with the release of TTS v0.1.3, the latest and greatest version of our text-to-speech engine.
Details below the fold; enjoy the newsletter!
📔Jupyter Notebooks
By Josh Meyer
Like Coqui? Like Python notebooks? You’re going to ❤️ this 😎.
We’ve added support for Jupyter notebooks in Coqui STT; so, you can train models easily from your browser! Check them out on Google Colab:
- Train a new model from scratch Open In Colab
- Train a model with transfer learning Open In Colab
If you want to run notebooks on your own machine, we also created a Coqui+Jupyter Dockerfile to make things easier.
🚀 Translators without Borders Deploying Coqui’s STT
By Josh Meyer
“Creating a world where knowledge knows no language barriers” - that’s the mission statement of Translators without Borders (TWB).
We’re happy to share that 🐸 STT is becoming a part of voice-enabled applications at TWB, starting with speech-to-text for Bengali! TWB trained a Bengali speech-to-text model, and they generously released the model in the Model Zoo.
This Bengali model is becoming part of TWB’s online glossary application for the Bengali language, which will enable faster search and translation.
Alp Öktem (Computational Linguist at TWB and co-founder of Col·lectivaT) trained the Bengali model using 🐸 STT in a little under 10 days using about 200 hours of open-source Bengali data.
Data formatting took up the bulk of development time, but once the data was formatted, training an STT model was “pretty straightforward” , says Öktem. Following the recommendations from a recent Coqui publication, Öktem found that fine-tuning from a pre-trained 🐸 STT model improved performance drastically compared to training from scratch.
Keep an eye out for new models and new voice-enabled applications from TWB!
🎥Coqui, Stars of Stage and (the Small) Screen
By Josh Meyer
Previously, we let you all know about our research collaboration with Harvard and Google. Now we have a 3-minute video overview of that project. The video was created with ❤️ by Mark Mazmuder from the Harvard Edge Computing Lab.
We hope you find the video accessible and interesting; and that it gives you a better idea of how we’re bringing speech technologies to all languages!
Out the Door, TTS v0.1.3
By Eren Gölge
This new version fixes important issues with the new TrainerAPI and introduces various performance improvements. Also, there is a new FullbandMelGAN vocoder released for the Thorsten German dataset. It achieves faster than real-time inference with the Tacotron2 DDC model.
Try out the German model:
pip install TTStts --model_name tts_models/de/thorsten/tacotron2-DCA \--text "Was geschehen ist geschehen, es ist geschichte."
Thanks to all the contributors who helped for this release:
🔬 Monthly TTS Papers
By Eren Gölge
We’ve also read some really interesting TTS papers this month. A few which were of particular interest:
- VITS: Conditional Variational Autoencoder with Adversarial Learning for End-to-End Text-to-Speech
- FastPitch: Parallel Text-to-speech with Pitch Prediction
- Efficient Deep Learning: A Survey on Making Deep Learning Models Smaller, Faster, and Better
Note: VITS and FastPitch models will be available in the next TTS release with pre-trained models. 🚀
👀 See our TTS-papers list for even more TTS papers!