voices.json speaker mapping for en_US-libritts_r-medium no longer matches published model
[Prepared by GPT 6 Astra]
The October 20, 2026 commit 5041a7a (“Regenerate example”) changed the en_US-libritts_r-medium entry in voices.json, but the published model configuration still contains the previous speaker mapping.
| Field | Published model config | Current voices.json |
|---|---|---|
| num_speakers | 904 | 2306 |
| speaker_id_map["17"] | 805 | 348 |
| speaker_id_map["104"] | absent | 1148 |
The preceding commit, b1cb611, has catalog speaker metadata matching the published configuration.
The file metadata also disagrees: voices.json lists the ONNX model as 81,472,438 bytes, but the download endpoint reports 78,580,914 bytes—the previous size.
This causes applications using the catalog’s speaker mapping to select different speakers or pass invalid IDs. A Read Aloud user on Windows 10 / Firefox reported voice 17 changing from male to female and voice 104 failing with:
Non-zero status code returned while running Gather node.
Name:'/emb_g/Gather'
indices element out of data bounds, idx=1148 must be within
the inclusive range [-904,903]
User report: https://github.com/ken107/read-aloud/issues/536
To verify, compare the en_US-libritts_r-medium entry in voices.json with en/en_US/libritts_r/medium/en_US-libritts_r-medium.onnx.json at commit 5041a7a.
Could you regenerate the catalog from the published model assets, or publish the matching model/configuration if this was an intended model update? A consistency check between catalog speaker mappings and model configurations would help prevent recurrence.