synthi tts

synthi-tts

Digitize your own voice to have it speak for you! Fully automated!

Notice by Niklas: I've taken this prototype and re-implemented it to make everything sound a lot better! I would like to thank everyone who helped me bring this idea to life, especially Moritz and Jackob who were responsible for the code in this repo!

---

Creating the voice model

• Speak a little! [audio.mp3]
• Transcribe what you've said [plain text - transcription.txt]
• use forced aligner (gentle) to match timecodes in audio and transcript
• extract phoneme slices (ffmpeg)

# prepare audio.mp3 and transcription.txt
# you need docker pre-installed
# this will generate an output.json inside the path specified.
./gentle-docker.sh <path> <audio.mp3> <transcription.txt>
python3 json-times-voice-to-samples.py <path/output.json> [output-folder]

Speech synthesis

• requires the voice model created prior to this step
• enter text
• convert to IPA-phonetics (espeak)
• intelligently concatenate voice samples from the model (ffmpeg)

# make sure that voice model is complete
# English is recommended
python3 synthi-tts.py "Your text here" -f <output-folder>

Use python3 synthi-tts.py --help for further help with arguments you may use.

TODO

• Matching and concatenation on syllable/word level for even better and more understandable synthesis
• Audio edits to remove glottal stops at the end of the short voice samples
• make this a lot faster!
• write an interface so it can be used as your voice system-wide
• graphical TTS utility