Training machine offline · last report yesterday
Voice training runs
Sanskrit text-to-speech. The goal is two voices of our own: one that chants verse by its metre, one that reads prose. Daily Sanskrit already chants with Vāgdhenu by Prathosh A P (IISc), credited there. No audio from these experiments is published.
Headline score: Letter error, bench · lower is better 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.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.
Every score on this track
- Letter error, bench (lower is better)
- 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.
- Letter error, whole bake-off (lower is better)
- Letter error over all 30 bake-off lines: verse, prose and hard cases (single words, clusters, visarga).
- Letter error, bake-off verses (lower is better)
- The same letter error on the bake-off's 14 verse lines only, so runs compare directly with the bake-off models.
- ā/a length ratio (higher is better)
- 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.
Live now
Progress
Lower is better on the headline.
Best Letter error, bench over time
Each dot is a scored run, placed at the day it ended; the line is the best score so far. Numbered lines are this track's milestones, listed below.
Milestones
- 9 Oct 2026 · TTS-E1: the first Sanskrit text-to-speech fine-tune starts on a cloud GPU
- 8 Oct 2026 · Voice bake-off: open Sanskrit text-to-speech models scored for letter error and vowel length
Reference models
Other models measured on the same sets, for comparison. Notes say how they were measured.
Method. The bake-off: 30 Sanskrit lines (14 verse, 6 prose, 10 hard cases), one take per model, transcribed by the Su-śrotā recogniser and compared letter by letter; vowel length by forced alignment of the known text.
| Model | Letter error, whole bake-off | Letter error, bake-off verses | ā/a length ratio |
|---|---|---|---|
| OmniVoice Lowest letter error in the bake-off, but its long vowels come out short. | 0.0408 | 0.0091 | 1.67 |
| Vāgdhenu By Prathosh A P (IISc). The only model with human-like vowel length; Daily Sanskrit chants with it, credited there. | 0.0424 | 0.0135 | 2.2 |
| Svara v1 Fewest errors by sound class (long vowels, visarga, aspirates). | 0.0528 | 0.0186 | 1.75 |
| Indic-Mio | 0.0585 | 0.0236 | 1.75 |
| vits_rasa_13 Small and fast enough for a phone. Trained on Rasa prose, so its long vowels are as short as that reader's. | 0.0952 | 0.0624 | 1.5 |
| Indic Parler-TTS Sampled output: takes of the same line vary by several points. | 0.0989 | 0.0468 | 1.5 |
| Indic-Speak Some takes ran on past the end of the text. | 0.1001 | 0.0537 | 1.8 |
| SPRING_F5 Drops words and the starts of words, and speaks too fast. | 0.1664 | 0.1746 | 1.33 |
| IndicF5 (base) The base model the Vāgdhenu recipe fine-tunes. Before that fine-tune it does not speak Sanskrit. | 0.6467 | 0.6492 | — |
| People reading Real speech, not a model: 60 recorded phrases measured the same way. The yardstick for vowel length. | — | — | 2.17 |
Voice runs
Published · 8 runs
| Data | |||||||||
|---|---|---|---|---|---|---|---|---|---|
| TTS-E4 tts_e4a_jaya |
9 Oct 2026 | IndicF5 | Archival recitation (non-Vedic), 3.3 h, 3.3 h | 4.09 | $5.32 | — | 0.0198 | 0.0302 | 2.14 |
| TTS-E3 tts_e3_v3 |
9 Oct 2026 | IndicF5 | Vāgdhenu + Rasa prose, 13.9 h | 5.33 | $6.92 | — | 0.007 | 0.0048 | 2.2 |
| TTS-E2 tts_e2_multi_b |
9 Oct 2026 | IndicF5 | Vāgdhenu + Su-śrotā scholars, 10.5 h | 4.83 | $6.29 | — | 0.0073 | 0.0075 | 2.2 |
| TTS-E2 try 1 Failed tts_e2_multi |
9 Oct 2026 | IndicF5 | Vāgdhenu + Su-śrotā scholars | 0.37 | $0.48 | — | — | — | — |
| TTS-E1 tts_e1_repro_d |
9 Oct 2026 | IndicF5 | Vāgdhenu chant, 5.2 h | 4.13 | $5.37 | — | 0.0072 | 0.0088 | 2.2 |
| TTS-E1 try 3 Failed tts_e1_repro_c |
9 Oct 2026 | IndicF5 | Vāgdhenu chant | 0.07 | $0.10 | — | — | — | — |
| TTS-E1 try 2 Failed tts_e1_repro_b |
9 Oct 2026 | IndicF5 | Vāgdhenu chant | 0.07 | $0.09 | — | — | — | — |
| TTS-E1 try 1 Failed tts_e1_repro |
9 Oct 2026 | IndicF5 | Vāgdhenu chant | 0.07 | $0.10 | — | — | — | — |
No runs match these filters. Clear the search or pick “All”.
Cost “—” means no cloud bill (our own GPU at home) or a cost that was not recorded; “Free” is free cloud compute. GPU-h is GPU hours.