Botium Speech Processing is a unified, developer-friendly API to the best available free and Open-Source Speech-To-Text and Text-To-Speech services.
Botium Speech Processing is a get-shit-done-style Open-Source software stack, the configuration options are rudimentary: it is highly opinionated about the included tools, just get the shit done.
- With Kaldi a reasonable speech recogniction performance is available with freely available data sources.
- MaryTTS is currently the best freely available speech synthesis software
- SoX is the swiss army-knife for audio file processing
While the included tools in most cases cannot compete with the big cloud-based products, for lots of applications the trade-off between price and quality is at least reasonable.
Read about the project history here
Some examples what you can do with this:
- Synthesize audio tracks for Youtube tutorials
- Build voice-enabled chatbot services (for example, IVR systems)
- Classification of audio file transcriptions
- Automated Testing of Voice services with Botium
- 8GB of RAM (accessible for Docker) and 40GB free HD space
- Internet connectivity
- docker
- docker-compose
Note: memory usage can be reduced if only one language is required - default configuration comes with two languages.
Clone or download this repository and start with docker-compose:
> docker-compose up -d
This will take some time to build.
Point your browser to http://127.0.0.1 to open the Swagger UI and browse/use the API definition.
This repository includes a reasonable default configuration:
- Use MaryTTS for TTS
- Use Kaldi for STT
- Use SoX for audio file conversion
- Languages included:
- German
- English
Configuration changes with environment variables. See comments in this file.
Recommendation: Do not change the .env file but create a .env.local file to overwrite the default settings. This will prevent troubles on future git pull
The environment variable BOTIUM_API_TOKENS contains a list of valid API Tokens accepted by the server (separated by whitespace or comma). The HTTP Header BOTIUM_API_TOKEN is validated on each call to the API.
For performance improvements, the result of the speech-to-text and text-to-speech calls are cached (by MD5-hash of audio or input text). To enforce reprocessing empty the cache directories:
- frontent/resources/.cache/stt
- frontent/resources/.cache/tts
Point your browser to http://127.0.0.1/ to open Swagger UI to try out the API.
Point your browser to http://127.0.0.1/dictate to open a rudimentary dictate.js-interface for testing speech recognition (for Kaldi only)
Attention: in Google Chrome this only works with services published as HTTPS, you will have to take of this yourself. For example, you could publish it via ngrok tunnel.
Point your browser to http://127.0.0.1/tts to open a MaryTTS interface for testing speech synthesis.
Available for Kaldi only
There are Websocket endpoints exposed for real-time audio decoding. Find the API description in the Kaldi GStreamer Server documentation.
The Websocket endpoints are:
- English: ws://127.0.0.1/stt_en/client/ws/speech
- German: ws://127.0.0.1/stt_de/client/ws/speech
Place audio files in these folders to receive the transript in the folder watcher/stt_output:
- watcher/stt_input_de
- watcher/stt_input_en
Place text files in these folders to receive the synthesized speech in the folder watcher/tss_output:
- watcher/tts_input_de
- watcher/tts_input_en
See swagger.json:
-
HTTP POST to /api/stt/{language} for Speech-To-Text
curl -X POST "http://127.0.0.1/api/stt/en" -H "Content-Type: audio/wav" -T sample.wav
-
HTTP GET to /api/tts/{language}?text=... for Text-To-Speech
curl -X GET "http://127.0.0.1/api/tts/en?text=hello%20world" -o tts.wav
-
HTTP POST to /api/convert/{profile} for audio file conversion
curl -X POST "http://127.0.0.1/api/convert/mp3tomonowav" -H "Content-Type: audio/mp3" -T sample.mp3 -o sample.wav
To be done: contribution guidelines.
We are open to any kind of contributions and are happy to discuss, review and merge pull requests.
This project is standing on the shoulders of giants.
- Kaldi GStreamer server and Docker images
- MaryTTS
- SVOX Pico Text-to-Speech
- Kaldi
- Kaldi Tuda Recipe
- Zamia Speech
- Deepspeech and Deepspeech German
- SoX
- dictate.js
- Optional start/end parameters for audio file conversion to trim an audio file by time codes formatted as mm:ss (01:32)
- Additional endpoint to calculate the Word Error Rate (Levenshtein Distance) between two texts
- When giving the hint-parameter with the expected text to the STT-endpoint, the Word Error Rate will be calculated and returned
- When multiple STT- or TTS-engines are configured, select the one to use with the stt or tts parameter (in combination with the Word Error Rate calculation useful for comparing performance of two engines)
- Using pre-trained models from Zamia Speech for speech recognition
- Using latest Kaldi build
- Added file system watcher to transcribe and synthesize audio files