mastodon - Link to source
New beta release of Sonata-for-NVDA, formaly known as Piper-for-NVDA
NVDA 2024.1 compatibility
Support for fast variants for Piper voices. These fast variants improves responsiveness significantly because they use streaming synthesis
Improvements to responsiveness and speed across the board
Release page:
github.com/mush42/sonata-nvda/…
Direct download link:
github.com/mush42/sonata-nvda/…

Release v3.0-beta.1 · mush42/sonata-nvda

What's new NVDA 2024.1 compatibility Changed the name to Sonata since we plan to support additional TTS models besides Piper in the future. Support for fast variants for Piper voices. These fast v...
GitHub

reshared this

in reply to Musharraf

mastodon - Link to source

Tamas G

yeah, and I love the speed improvement it gives to some of those! Especially Mac Alex :) I did notice though that the faster models do a way better job at not cutting the starting fragments of words (older voices can still slightly do this.) I see that encode and decode are now two models so feels like splitting older voices into the new format is less straight forward without a full retrain, but I might be entirely wrong.
in reply to Tamas G

mastodon - Link to source

Musharraf

@Tamasg
If you have the original checkpoint, you can convert it to the new format.
Take a look at this script, which I used to export Piper's checkpoints:
github.com/mush42/piper-rt-mak…

I need to update the docs, and add a section on training voices.


piper-rt-maker/tasks.py at main · mush42/piper-rt-maker

Export and package Piper checkpoints as RT voices in ONNX format - mush42/piper-rt-maker
GitHub
@Tamas G

Tamas G reshared this.

in reply to Musharraf

mastodon - Link to source

Tamas G

oooh this is so superb. So I could still continue to train with the existing notebook and eventually just convert the checkpoint CKPT file into the new RT voice, that's going to be amazing. At least it unblocks my work though with sourcing higher quality data and transcripts for the voices I've developed, so just already knowing that is huge help. I could point the variables to URLs in my drive folder for the checkpoint, then convert that way with the script you referenced.
in reply to Tamas G

mastodon - Link to source

Musharraf

@Tamasg
Here're the steps to convert the checkpoint to fast format:
# Clone piper fork containing export code
git clone github.com/mush42/piper
cd ./piper
# Checkout streaming branch
git checkout streaming
cd ./piper/src/python
pip3 install -r requirements.txt
# Upgrade torch
pip3 install --upgrade torch pytorch-lightning onnx
source ./build_monotonic_align.sh
# Export. Edit paths
python3 -m piper_train.export_onnx_streaming --debug [checkpoint path] [export directory]

GitHub - mush42/piper: A fast, local neural text to speech system

A fast, local neural text to speech system. Contribute to mush42/piper development by creating an account on GitHub.
GitHub
@Tamas G
in reply to Musharraf

mastodon - Link to source

Tamas G

ok. I'm not sure if this is good or bad, and I hope it doesn't insult your script code too much by gutting it, but I made a notebook file of your steps: eurpod.com/Export_piper_voice_… - hopefully this looks correct, it restructures it a bit to take a path to a file stored in drive after mounting. I'll try it in a little bit and see if it's a spectacular fail or not :D (minor correction to made for running the .sh file since we can't use virtual environments.)
This entry was edited (1 year ago)
in reply to Pratik Patel

mastodon - Link to source

Musharraf

@ppatel
It seams like the server is not running.
Are you running NVDA on a 32-bit/ARM-64 machine. Sonata only works on 64-bit versions of Windows.
Otherwise, check if the server generated any logs in the following file path:
[NVDA config directory]\sonata\logs\sonata-grpc.log
If not, try running the following binary from a cmd window and report the output:
[NVDA config directory]\addons\sonata_neural_voices\synthDrivers\sonata_neural_voices\bin\sonata-grpc.exe
@Pratik Patel
in reply to Musharraf

friendica (DFRN) - Link to source

Peter Vágner

Dear @Musharraf :verified: Can you please give me a hint how do I build the file sonata-grpc.exe is it a result of building github.com/mush42/sonata ? We are working on a slovak human sounding voice with friends and I am tweaking corresponding espeak-data along the way, so until I manage to get these pushed and merged to espeak-ng I imagine my best bet is rebuilding the addon with all the resources locally.
Thanks for all the fantastic work you are putting into this.
@Musharraf
in reply to Peter Vágner

mastodon - Link to source

Musharraf

@pvagner
Here's how to build the sonata-grpc binary:
git clone github.com/mush42/sonata
cd ./sonata/sonata-grpc
# With Rust installed
cargo build --release

GitHub - mush42/sonata: A cross-platform engine for neural TTS models.

A cross-platform engine for neural TTS models. Contribute to mush42/sonata development by creating an account on GitHub.
GitHub
@Peter Vágner
Unknown parent

mastodon - Link to source

Tamas G

I did train it up to 6000 epochs over night. Even though like 3K is recommended for fine-tuned models by Piper, but I gave it a lot more. I do feel like it's better, but could be placibo. If you have the energy you can re-install and it should overwrite: eurpod.com/en_us-MacAlex+RT_me… (one odd thing this and the other voice do is say numbers that are short but with the hundredth digit mark oddly.)
This entry was edited (1 year ago)
in reply to Andre Louis

mastodon - Link to source

Tamas G

@FreakyFwoof @fireborn @TomGrant91 the old voices in the newer add-on will still give a similar speed improvement. I guess the one thing that disappointed me slightly - even if you import the older Keynote or Votrax voice files into that Sonata as voices, they will still have quite good speeds almost to the same degree for me. I wonder if that's just me though.
@aaron @Tom Grant @Andre Louis
in reply to Tamas G

mastodon - Link to source

JamminJerry

@Tamasg @FreakyFwoof @fireborn @TomGrant91 I need to figure out why the sonata neural voices won't work at all for me anymore. when I try to switch to it, it says it can't load the sonata neural voices. I was running the first beta, and that was when it worked for a couple days then stopped working all of a sudden. I have completely uninstalled the voices, and then the addon, and installed the beta 2 of the addon, and some voices, but it still won't load it at all. I have even reset the computer thinking something might have just been messed up, but not even that has helped.
@aaron @Tamas G @Tom Grant @Andre Louis
⇧