Piper keeps downloading already downloaded models.

[vindolin]

Running the command: echo 'Welcome to the world of speech synthesis!' | piper --model en_US-danny-low --output-raw | aplay -r 22050 -f S16_LE -t raw works on my RPi4, but when I repeat the same command. it outputs the warning: WARNING:piper.download:Wrong size (expected=5953, actual=4166) for /home/vindolin/projects/piper/en_US-danny-low.onnx.json and repeats downloading the same files before playing the output.

Any ideas?

[vindolin]

I dug around a bit and discovered that the file sizes and md5 hashes in my /home/tulpe/.virtualenvs/piper/lib/python3.9/site-packages/piper/voices.json differed from the ones on https://huggingface.co/rhasspy/piper-voices

[Tristan-Hopkins]

I've noticed the same thing, however, it seems harmless. a simple fix is to run it like this instead - echo 'This is a test.' | \ piper -m /path/to/model.onnx --output_file test.wav

[ichthyosaurus]

Have you tried running piper --update-voices -m <any model> to fetch the latest voices.json? Piper prefers the voices.json found in your data dir over the bundled one.

I fixed a related bug in #367.

[wentbackward]

Have you tried running piper --update-voices -m <any model> to fetch the latest voices.json? Piper prefers the voices.json found in your data dir over the bundled one.

I fixed a related bug in #367.

This works for me.

[jn64]

piper-tts 1.2.0 via pip
Python 3.11
Fedora 38

I can repro the original report (without --data-dirs, so #367 doesn't apply). On first run, piper downloads a model, then on subsequent runs, the $model.onnx.json file is re-downloaded every time due to wrong size. --update-voices fixes it.

From what I understand, the problem is:

1. Without voices.json in the cwd (or $DATA_DIR? not clear to me), piper falls back to the bundled voices.json which may be outdated. E.g. in 1.2.0, the size of en_US-lessac-medium.onnx.json was 7010 B, but it's 4885 B in the current file on huggingface (btw I don't know how/why this file differs from this repo)

2. If piper is called with a model name (as opposed to a path to a model file), it will check voices.json. If the local files matching that name do not match the size (or checksum?) in voices.json, they are re-downloaded on every use.

Although the $model.onnx.json files are tiny, the constant re-downloading is unnecessary, and adds unexpected latency to the speech synthesis.

Suggestions:

1. Check and update voices.json on first run (i.e. do --update-voices by default if voices.json doesn't exist in cwd). Piper already downloads the model files on first run, and voices.json size is negligible in comparison. This does not prevent voices.json eventually becoming outdated again, but at least out of the box, the user is not unexpectedly downloading files repeatedly.

2. Add some file modified time check to reduce superfluous downloads. Even people who want the model to be auto-updated probably don't need it to be updated on every single use.

3. Alternatively, default to no network access and always favour local files. Add an option to update both voices.json and models (e.g. --update, replacing --update-voices). If any files don't match voices.json there should be a warning printed but no network access without the option. There could be a prompt to download if the specified models don't exist locally, to make it easier for first time users.

   It can be achieved by specifying a path to the model file instead of model name, but the current behaviour (even if this issue is fixed) is unexpected since one of the draws of piper is that it is "local".