Pretrained playing-technique recognition models for ipt~, pipo.ipt and any host built on libipt.
Models are TorchScript (.ts) files trained and exported with ipt_recognition.
| Model | Instrument | Classes | Sample rate | Latency floor | Card |
|---|---|---|---|---|---|
eguitar_ircam.ts |
Electric guitar | 14 | 8 kHz | 896 ms | card |
flute_ircam.ts |
Flute | 11 | 44.1 kHz | 333 ms | card |
trumpet_ircam.ts |
Trumpet | 14 | 44.1 kHz | 333 ms | card |
trumpet_harmon_ircam.ts |
Trumpet, harmon mute | 14 | 44.1 kHz | 333 ms | card |
Each card lists the classes in output order, the model's specifications and its SHA-256.
Put the .ts file in Max's search path (or give an absolute path), then:
ipt~ flute_ircam.ts
pipo~ ipt @ipt.model flute_ircam.ts
The host audio is resampled to the model's sample rate by libipt, so any Max sample rate works.
Models are hosted on Hugging Face: nbrochec/ipt_models. The links in the table above always point to the latest version; previous versions stay available in the repository history. From the command line:
hf download nbrochec/ipt_models flute_ircam.ts --local-dir .
A SHA256SUMS file is provided alongside the models: shasum -a 256 -c SHA256SUMS.
cards/ one Markdown card per model
This project is released under a CC-BY-NC-4.0 license.
Please write to nicolas.brochec[at]ircam.fr for any questions.