Audio-to-Audio
PEFT
Safetensors
Moshi
Turkish
Hindi
speech-to-speech-translation
simultaneous-translation
mimi
lora
tpu
turkish
hindi
Eval Results (legacy)
Instructions to use tiny-aya-translate/tr-hi-s2st-v0.3 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- PEFT
How to use tiny-aya-translate/tr-hi-s2st-v0.3 with PEFT:
Task type is invalid.
- Moshi
How to use tiny-aya-translate/tr-hi-s2st-v0.3 with Moshi:
# pip install moshi # Run the interactive web server python -m moshi.server --hf-repo "tiny-aya-translate/tr-hi-s2st-v0.3" # Then open https://localhost:8998 in your browser
# pip install moshi import torch from moshi.models import loaders # Load checkpoint info from HuggingFace checkpoint = loaders.CheckpointInfo.from_hf_repo("tiny-aya-translate/tr-hi-s2st-v0.3") # Load the Mimi audio codec mimi = checkpoint.get_mimi(device="cuda") mimi.set_num_codebooks(8) # Encode audio (24kHz, mono) wav = torch.randn(1, 1, 24000 * 10) # [batch, channels, samples] with torch.no_grad(): codes = mimi.encode(wav.cuda()) decoded = mimi.decode(codes) - Notebooks
- Google Colab
- Kaggle
README: interim 12-point ladder disclosure (pre-release)
Browse files
README.md
ADDED
|
@@ -0,0 +1,19 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: cc-by-nc-4.0
|
| 3 |
+
---
|
| 4 |
+
|
| 5 |
+
# tr-hi-s2st-v0.3 (private, pre-release)
|
| 6 |
+
|
| 7 |
+
Checkpoint suite for the TinyAya v0.3 Turkish↔Hindi speech-to-speech
|
| 8 |
+
translation run (`v0.3-long-horizon-mh-r2`, completed by early stop at step
|
| 9 |
+
65,250; best `val/composite` 2.9048 @ step 62,750).
|
| 10 |
+
|
| 11 |
+
**Interim checkpoint ladder (private-storage limits):** this repo currently
|
| 12 |
+
carries a 12-point trajectory — branches `best` (step 62,750), `step-65250`
|
| 13 |
+
(final), and `step-{6000,12000,…,60000}` (every 6,000). The full Pythia-style
|
| 14 |
+
per-1,000 suite (78 checkpoints, plus log-spaced early steps) exists in
|
| 15 |
+
archival storage and will be published here in full at release.
|
| 16 |
+
|
| 17 |
+
Load a point: `from_pretrained("tiny-aya-translate/tr-hi-s2st-v0.3", revision="step-60000")`.
|
| 18 |
+
Audio demos: `samples/step_N/`. Training log: `logs/` (deduped; see header note).
|
| 19 |
+
Eval numbers land with the release (see the repo's eval harness: `scripts/eval_release.py`).
|