tts_e4a_jaya
Fine-tune of IndicF5 on Archival recitation (non-Vedic), 3.3 h, 3.3 h, trained on A100 40GB (Google Cloud, spot). Started 9 Oct 2026, ended 10 Oct 2026.
Hypothesis
A chant voice learned from about 3 hours of archival recitation, used with the family's permission: can a few hours of older, band-limited recordings make a clear chant voice?
What we learned
Bench letter error 0.0198, ā/a ratio 2.14; write-up pending.
Scores
Lower is better on the headline; change against the parent run, TTS-E2.
| Test set | TTS-E4 | TTS-E2 (parent) | Change |
|---|---|---|---|
| Letter error, bench headline Each render is transcribed by the Su-śrotā Sanskrit speech recogniser and compared letter by letter with its text: 0.04 means 4% of letters. Mean over the bench: 30 held-out chanted verses and the bake-off's 14 verse lines. The verses' own human recordings score 0.011 on it. |
0.0198 | 0.0073 | +171.2% |
| ā/a length ratio How much longer a long ā lasts than a short a, by forced alignment of the known text (median, inside words). Sanskrit wants about 2: people reading measure 2.17 here; most models say long vowels too short. |
2.14 | 2.2 | −2.7% |
| Letter error, bake-off verses The same letter error on the bake-off's 14 verse lines only, so runs compare directly with the bake-off models. |
0.0302 | 0.0075 | +302.7% |
Letter error rate: the share of letters that come out wrong, as a fraction (0.0096 means 0.96% of letters): for OCR, letters read off a page; for a voice, letters a speech recogniser hears in its audio. Lower is better.
Curves
Training loss
The flow-matching loss the speech model trains on (how far its predicted sound is from the recording's), averaged over each logged stretch, by step. It falls slowly and noisily; the scores come from the bench at the end.
Model
- Preset
- IndicF5
- Parameters
- —
- Layers · heads · width
- 22 · 16 · 1,024
Data
- Recordings
- Archival recitation (non-Vedic), 3.3 h, 3.3 h
- Audio published
- None: these voices stay internal
Compute
- GPU
- A100 40GB
- Where
- Google Cloud (spot)
- Steps
- 15,000 / 15,000
- GPU hours
- 4.09
- Cost
- $5.32
- Spot restarts
- 2
Lineage
gcpa100spotindicf5