fleurs-tr-hi-mimi-encoded

Mimi-encoded Turkish↔Hindi parallel speech pairs for TinyAya Stage 2 speech-to-speech translation training.

Contents

• encoded/*.pt — 9212 Mimi-encoded audio pairs (kyutai/mimi, 8 codebooks, 12.5 Hz, 24 kHz). Each file keys: pair_id, src_lang, tgt_lang, src_text, tgt_text, src_codes[8, T_src], tgt_codes[8, T_tgt].
• encoded/*.alignments.json — 18424 Whisper word-level alignment sidecars (.src.alignments.json / .tgt.alignments.json).
• splits/train.jsonl — 8283 rows (90%)
• splits/val.jsonl — 929 rows (10%)

Provenance

Sourced from tiny-aya-translate/fleurs-tr-hi-parallel-speech (accepted.jsonl sha256 8a67ddada1f13e5ec11d2bd9ee78a4acd1d5543c665a2cba04556a3c919043ad). Mimi config: kyutai/mimi, mimi_num_codebooks=8, output_sample_rate=24000. Splits built by scripts/make_splits.py (seed=42, 90/10 keyed on FLEURS sentence_id — zero overlap).

⚠️ Acoustic-shift evaluation set, not held-out text. An audit found 200/200 sampled texts also occur in the training corpus (FLORES ⊂ FLEURS), so this measures robustness to real human speech only — never describe results on it as held-out text.

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

Model	tr-hi-s2st-v0.3
Text	tr-hi-parallel-text
Speech	tr-hi-parallel-speech-v2 · -v3
Encoded	tr-hi-mimi-encoded
Eval sets	fleurs-tr-hi-mimi-encoded · lahaja-eval · cv-tr-eval

Code

repo	what it does
model	Stage-2 training, evaluation harness and TPU launch tooling

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).