README.md

metadata

title: CantusSVS
emoji: 🕊️
colorFrom: gray
colorTo: blue
sdk: streamlit
sdk_version: 1.32.2
app_file: app.py
pinned: false

CantusSVS

Table of Contents

• About CantusSVS
• Preparing Your Input
• FAQ

---

About CantusSVS

CantusSVS is a singing voice synthesis tool that automatically generates audio playback for the Latin chants in Cantus. For training and inferencing, we use DiffSinger, a diffusion-based singing voice synthesis model described in the paper below:

DiffSinger: Singing Voice Synthesis via Shallow Diffusion Mechanism

Liu, Jinglin, Chengxi Li, Yi Ren, Feiyang Chen, and Zhou Zhao. 2022. "Diffsinger: Singing Voice Synthesis via Shallow Diffusion Mechanism." In Proceedings of the AAAI Conference on Artificial Intelligence 36 10: 11020–11028. https://arxiv.org/abs/2105.02446.

Training was done using Cedar, a cluster provided by the Digital Research Alliance of Canada. To set up training locally, follow this tutorial by tigermeat.

For general help training and creating a dataset, this tutorial by PixPrucer is an excellent guide. For help, join the DiffSinger Discord server.

The dataset used for this project was built using Adventus: Dominica prima adventus Domini, the first track from Psallentes' album Salzinnes Saints. Psallentes is a Belgian women's chorus that specializes in Late Medieval and Renaissance music. Salzinnes Saints is an album of music from the Salzinnes Antiphonal, a mid-sixteenth century choirbook with the music and text for the Liturgy of the Hours.

---

Preparing Your Input

• Most commercial music composition software can export .mei files. MuseScore 4 is free to use.
• Input format must be .mei (Music Encoding Initiative).
• Only monophonic scores are supported (one staff, one voice).
• Lyrics must be embedded in the MEI file and aligned with the notes.

Validation tool:

python scripts/validate_mei.py your_song.mei

---

FAQ

Q: Can I synthesize polyphonic (multi-voice) chants?
A: No, only monophonic scores are supported currently. However, in the future, polyphonic chants could be synthesized by layering multiple monophonic voices.

Q: Can I change the voice timbre?
A: In the webapp, only the provided pre-trained model is available. However, DiffSinger will learn the timbre of the input dataset so if you train your own model, you can control the timbre that way.

---