Tacotron 2 synthesis wrapper with syntax based on/similar to notaistreamer's Forsen TTS donation system.
+1 BTC +10 社会信用
Required:
- Python >= 3.8
- Pytorch 1
- sox (for sound filters)
Then install the rest (in a virtual env)
pip install -r requirements.txt
Tacotron models I have pretrained can be downloaded from here (338M per model).
You can instead run/edit download-models.py which handles naming
automatically (otherwise you would have to type udisen2-114: instead of
udisen:)
If you want to use your own models, place them in the models directory.
Run main.py to launch interactive prompt, or with --stdin to read from
standard input.
To use CPU for inference: ./main.py --cpu
You can also set the number of threads: ./main.py --cpu --threads 2
Using the syntax speaker: ...rest of message will use the file
models/speaker.
./main.py --stdin <<< "peroni: are you happy? to steal my voice? fakin tresh."
mpv audio.wav
./main.py --stdin < example-texts/poem.txt
mpv audio.wav
./main.py --stdin <<< "{fembaj.py} velcuz: remember to subscribe to my only \
fans, fifty percent off. only fans slash velcuz, limited time offer."
mpv audio.wav
The filters are the same as notaistreamer's filters (with the exception that every filter can stack as much as you want).
{1} room echo
{2} hall echo
{3} outside echo
{4} pitch down
{5} pitch up
{6} telephone
{7} muffled
{8} quieter
{9} ghost
{10} chorus
{11} slow down
{12} speed up
{.} stop previous filter
You can write a script in the sound-filters directory that takes the first
argument as the input audio file and the second argument as the destination
file. Scripts that return a non-zero exit code will have no filter applied.
See sound-filters/fembaj.py and sound-filters/fembaj.sh for examples.
Sound effects are mostly the same, except that sound names can be
anything; [abc] will use the file sound-effects/abc.wav.
notaistreamer's sound effects from 101soundboard can be downloaded from here (5.6M), which should save a lot of time from having to click every single sound effect and wait for an advertisement.
You can host an HTTP server with flask, which allows you to synthesize audio
on another machine or automate synthesis. This can be useful when stream
sniping, for example.
The HTTP server will accept POST requests at /. It will take data as plain
text and synthesize audio from it with flask's send_file.
On the machine you will be hosting the server on:
flask run
On your local machine that will make requests:
curl -X POST http://{IP_ADDRESS_OF_SERVER}:{PORT} \
-d "udisen: hello today will show how can stream snipe with text to speech" \
--output audio.wav
This takes input from rofi and synthesizes speech then plays it back.
rofi -dmenu | curl -X POST http://{IP_ADDRESS}:{PORT} \
-d @- \
--output /tmp/tts.wav && mpv /tmp/tts.wav