valenti/spark-vibevoice
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Update Discord support link with additional resource

Browse Source 
Added a link to the VibeVoice-finetuning repository for reference.
This commit is contained in:
mrfakename
2025-09-22 18:23:36 -07:00
committed by GitHub
parent f22fc38448
commit f1b697c157
1 changed files with 2 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files
FINETUNING.md
+2 -2
View File
@@ -5,7 +5,7 @@
More instructions coming soon.
Join the [Discord](https://discord.gg/ZDEYTTRxWG) for support.
Join the [Discord](https://discord.gg/ZDEYTTRxWG) for support. Also take a look at [voicepowered-ai/VibeVoice-finetuning](https://github.com/voicepowered-ai/VibeVoice-finetuning).
Example script:
@@ -47,4 +47,4 @@ python -m vibevoice.finetune.train_vibevoice \
- This is an unofficial finetuning implementation, it has not been validated by the original authors.
- The `voice_prompts_column_name` parameter is currently set to `audio` in the example above, which means the same audio file is used for both training data and voice prompts. This is appropriate when you don't have separate voice prompt files. However, if your dataset includes dedicated voice prompt files (short audio clips that capture the target speaker's voice characteristics), you should specify a different column name that contains these separate voice prompt files. For podcast-style training (once it is supported), the model typically uses the first utterance from each speaker within the podcast episode as the voice prompt, meaning the voice prompt is extracted from the beginning of the same audio file used for training.
- The dataset text/transcript must be in the format of "Speaker X: text", even if there is only one speaker. Example: `Speaker 1: Hello, how are you?`
- The default dataset is the Jenny (Dioco) dataset. This is a small dataset for testing purposes and each segment is only a few seconds long. The model may struggle to generate long audio with this dataset.
- The default dataset is the Jenny (Dioco) dataset. This is a small dataset for testing purposes and each segment is only a few seconds long. The model may struggle to generate long audio with this dataset.