meeami-tts checkpoints
Inference checkpoints for meeami-tts, a time-aligned US-English TTS. Give it a sentence and a recording of that sentence by a non-US speaker. It says the sentence in the LJSpeech voice with a US accent, on the reference's word timing, and the output is exactly as long as the reference.
Code, report and decision log: https://github.com/rrtchaganti/meeami-tts (docs/REPORT.md). The
checkpoints are fetched by make models (scripts/fetch_models.py), which checks every file
against MANIFEST.sha256; the test outputs by make outputs (outputs/test/MANIFEST.sha256).
Files
| Path | What it is | Size |
|---|---|---|
vits_mfa/m1_pos_diph/infer.pt |
Main system. VITS fine-tuned with MFA durations instead of MAS, plus a position-in-phone embedding and diphthong split | 142 MB |
vits_mfa/m1/infer.pt |
M1: VITS fine-tuned with MFA durations | 141 MB |
vits_mfa/m1_pos/infer.pt |
M1 + position-in-phone embedding | 141 MB |
vits_mfa/abl_a_pos/infer.pt |
Ablation A: m1_pos + 40% pace-augmented clips (LJ pace statistics) | 141 MB |
vits_mfa/abl_b_pos/infer.pt |
Ablation B: m1_pos + 40% pace-augmented clips (dev-accent pace statistics) | 141 MB |
baselines/vits/pretrained_ljs.pth |
V0 baseline: unmodified mirror of the official pretrained LJSpeech VITS (jaywalnut310/vits) | 146 MB |
baselines/fastspeech2/900000.pth.tar, .../hifigan/generator_LJSpeech.pth.tar |
M0 baseline: unmodified mirror of ming024/FastSpeech2's LJSpeech checkpoint and HiFi-GAN | 474 MB |
| lexicon/lexicon.txt | The frozen pronunciation lexicon (the only pronunciation source) | 0.4 MB |
| outputs/test/<system>.tar (7) | Every evaluated held-out test output: 1,248 utterances per system, each wav/<id>/out.wav + timings.json, the MFA TextGrid measured on it (align_out/), and per_utt.csv / summary.json / scores.jsonl. The exact files behind every reported number | 207 MB each |
Each infer.pt holds only what inference reads: the generator weights (net_g), the training
config and the step. The discriminators and optimizer states are dropped, which takes the file
from 956 MB to 141 MB. Loading it gives bit-identical weights and waveforms to the full training
checkpoint (scripts/export_infer.py --verify), and source.sha256 names that checkpoint.
Results on held-out accents (Arabic, Korean, Vietnamese; 12 speakers, 1,248 references)
| System | Word-boundary MAE | Pause F1 | WER | US prob | UTMOS |
|---|---|---|---|---|---|
| V0 (original VITS, MAS) | 40.5 ms | 0.815 | 2.1% | 0.649 | 4.31 |
| M0 (FastSpeech2) | 31.9 ms | 0.726 | 9.3% | 0.746 | 3.96 |
| M1 | 24.2 ms | 0.822 | 3.8% | 0.734 | 4.19 |
| m1_pos_diph (main) | 23.9 ms | 0.835 | 2.9% | 0.762 | 4.20 |
Licences and training data
m1,m1_pos,m1_pos_diph,abl_a_pos: fine-tuned from the MIT-licensed pretrained VITS on LJSpeech 1.1 (public domain) only;abl_a_posadds time-stretched LJSpeech copies.abl_b_pos: the time-stretched LJSpeech copies follow pace statistics measured on L2-ARCTIC dev speakers (CC BY-NC 4.0). No L2-ARCTIC audio was used in training. Use it for non-commercial research only, to be safe.baselines/: unmodified third-party weights, mirrored with their MIT licences (seeLICENSES.md).outputs/test/: synthesized speech only (the LJSpeech voice), but its texts and word timings come from L2-ARCTIC test recordings (CC BY-NC 4.0): non-commercial research only.- No L2-ARCTIC audio is stored in this repository. Each
timings.jsonnames the reference file it used.