tr-hi-s2st-v0.2

TinyAya — Turkish⇄Hindi Speech-to-Speech Translation (v0.2)

⚠️ Research preview — this checkpoint overfit. Released for transparency and to document the full training trajectory. The recommended weights are best_by_val (step 1,000), not the final step-15,000 checkpoint. See Evaluation & the overfitting story.

⚠️ Dataset disclosure — honest correction

This checkpoint was trained on tiny-aya-translate/fleurs-tr-hi-mimi-encoded — Mimi-encoded FLEURS Turkish↔Hindi read speech (~27% real FLEURS audio + ~73% multi-voice TTS over FLEURS text; ≈8.3k train / 929 val).

This was not the dataset we intended to train on. The project's synthetic data pipeline — FLORES + OPUS-100 + machine-translated conversational text, rendered with multi-voice TTS into tiny-aya-translate/tr-hi-mimi-encoded (~1.3M clips) — is the corpus our accompanying write-up describes. The training launcher's HF_DATASET default pointed at the fleurs- sibling repo, so v0.2 silently trained on FLEURS read speech instead of the synthetic conversational corpus. We are disclosing this openly rather than quietly re-labeling the run.

v0.3 corrects the data source (→ tr-hi-mimi-encoded) together with the codebase fixes (parallel-stream collator, regularization, deep-codebook weighting). See tr-hi-s2st-v0.3.

Moshi-style simultaneous speech-to-speech translation for Turkish ⇄ Hindi: a LoRA-fine-tuned Cohere2 backbone fused with a frozen Moshi depth decoder, operating on Mimi audio codes in a parallel two-stream format.

• Developed by: tiny-aya-translate
• Funded by: Google TPU Research Cloud (TRC) — see Acknowledgements
• Model type: parallel two-stream S2ST (Cohere2 + LoRA → CB0; frozen Moshi depth decoder → CB1–7)
• Languages: Turkish (tr), Hindi (hi)
• License: CC-BY-NC-4.0 — non-commercial, inherited from the base model (training code is Apache-2.0; see the caveat below)
• Finetuned from: CohereLabs/tiny-aya-base (+ frozen Moshi / Mimi from Kyutai)

Provenance: sweep → production run

This recipe was selected by a proxy-first W&B hyperparameter sweep (8 Bayesian

• hyperband trials), then trained to 15k steps on a single TPU v6e-8.

• 📊 Sweep: https://wandb.ai/cataluna84/tinyaya-stage2-tpu/sweeps/9ba8h0ho
• 📈 Training run: https://wandb.ai/cataluna84/tinyaya-stage2-tpu/runs/t1840nkd
• Code: the training repo (PR #8)

Training procedure

Evaluation & the overfitting story

The run was mechanically flawless but the recipe overfit: validation loss bottomed at step 1,000 and rose monotonically while train loss kept falling.

Final per-codebook val accuracy: cb0 13.8%, cb1 3.9%, cb2 1.9%, cb3 1.4%, cb4 0.8%, cb5 0.8%, cb6 0.6%, cb7 0.5%. The text stream effectively memorized (train text loss → 0.39); audio is the bottleneck. Likely cause: the proxy sweep optimized short-horizon val/audio_loss, which favored high capacity (lora_r=64) that then memorized the train set over 15k steps.

Downstream metrics (to be measured)

No speech-quality metrics have been computed yet. Planned evaluation, with the checkpoint to use = best_by_val (step 1,000):

Reporting protocol (when filled): ASR backend + version, decoding (greedy/beam), seed, text normalization, and number of eval samples will be stated.

Checkpoints in this repo

keep_last_n rotation during the run means only these survived (steps 2,000–12,000 were not retained — fixed for future runs):

Each checkpoint contains the composite components — depth_decoder.pt, text_embed.pt, audio_heads.pt, model_audio_embed.pt, projection.pt, metadata.json — plus the LoRA adapter under peft_adapter/ (adapter_model.safetensors + adapter_config.json). train_15k.log is the full training log.

Intended uses & limitations

• Intended: research, reproduction, and studying the TR⇄HI S2ST + overfitting trajectory. A teaching artifact for sweep→run→diagnosis.
• Out of scope / limitations: not production-ready — overfit, low per-codebook accuracy, no human eval, narrow domain (FLEURS read speech). May hallucinate or produce low-naturalness audio. Do not deploy for real translation without re-training (see below).

How to use

This is a composite model (custom architecture), not a drop-in transformers pipeline. Load via the training repo's src/model/composite.py: base Cohere2 backbone + the LoRA adapter (peft_adapter/) + the .pt components, then decode Mimi codes to audio. See the repo README for the loading + inference path. Use the best_by_val folder.

Recommended next run (fixing the overfit)

Lower capacity + regularization + early stopping: lora_r 16–32, higher weight_decay/dropout, stop at ~1–2k steps (or use best-N checkpoint averaging), and/or more/augmented data.

Bias, risks & limitations

Trained on FLEURS (read speech, limited speakers/domains); quality and fairness across dialects, accents, code-switching, and spontaneous speech are untested. Speech translation can mistranslate, omit, or fabricate content — outputs must not be relied upon for high-stakes communication.

License caveat (important)

These weights are CC-BY-NC-4.0 — non-commercial use only. They are LoRA derivatives of CohereLabs/tiny-aya-base, which is CC-BY-NC-4.0, so the non-commercial term is inherited; Apache-2.0 was not ours to grant over them and an earlier version of this card said so in error. The authors' separate training/eval code remains Apache-2.0, and the Moshi/Mimi components are CC-BY-4.0.

Acknowledgements

This model was trained on Cloud TPU v6e-8 hardware generously provided by Google's TPU Research Cloud (TRC) program. We thank the TRC team for supporting this research.

Citation

@misc{tinyaya_tr_hi_s2st_v0_2,
  title  = {TinyAya: Turkish-Hindi Speech-to-Speech Translation (v0.2)},
  author = {tiny-aya-translate},
  year   = {2026},
  note   = {Cohere2 + frozen Moshi depth decoder, LoRA, trained on Google TRC TPU v6e-8},
  url    = {https://huggingface.co/tiny-aya-translate/tr-hi-s2st-v0.2}
}

Where this sits

The v0.3 speech-to-speech pipeline, end to end:

tr-hi-parallel-text          text triples (en pivot -> tr / hi)
        |  TTS
tr-hi-parallel-speech-v2     synthetic speech + QC signals
        |  Mimi encode
tr-hi-mimi-encoded           8-codebook tokens + word alignments
        |  Stage-2 training
tr-hi-s2st-v0.3              the released model

Code

Project

TinyAya Stage 2 — Turkish⇄Hindi speech-to-speech translation with a text inner-monologue: a LoRA-adapted Cohere2 backbone driving a frozen Moshi depth decoder over Mimi codes.

The v0.3 run covered 76,250 steps / 2.07 epochs on a Cloud TPU v6e-16 (best val composite 2.8199 @ step 76,000). Read honestly: the text inner-monologue learns to translate (free-run chrF++ ~25.7 / 25.1), while intelligible audio synthesis remains the frontier (ASR-chrF++ 3.7 / 9.6 against a 92.1 / 86.6 ground-truth-audio ceiling) — bounded by the frozen depth decoder, not by translation understanding.

• Results: v0.3 evaluation report
• Training run: W&B xzcb60bl · emergence report
• Blog: Adapting Moshi for Low-Resource Speech Translation

Compute for the v0.3 run was provided by Google's TPU Research Cloud (TRC).