Datasets:

tiny-aya-translate
/

hinglish-casual

dataset_info:
  features:
    - name: id
      dtype: string
    - name: utterance
      dtype: string
    - name: utterance_latin
      dtype: string
    - name: style_metadata
      dtype: string
    - name: speaker
      dtype: string
    - name: duration
      dtype: float32
    - name: audio
      dtype: audio
    - name: word_count
      dtype: int32
  splits:
    - name: train
      num_bytes: 31017252549
      num_examples: 33275
  download_size: 29048443091
  dataset_size: 31017252549
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
language:
  - hi
  - en
license: other
license_name: see-upstream
task_categories:
  - automatic-speech-recognition
tags:
  - hinglish
  - code-switching
  - casual-speech
pretty_name: Hinglish Casual Speech

Hinglish Casual Speech

33,275 casual Hindi-English code-switched utterances (~31 GB) with audio, transcripts in both Devanagari and Latin script (utterance / utterance_latin), speaker ids, style metadata and durations. Full schema is in the YAML header above.

Collected during the TinyAya programme to probe code-switched speech, which neither the FLORES-derived text nor the TTS corpora cover. It is not part of the v0.3 Stage-2 training set — that is tr-hi-mimi-encoded.

from datasets import load_dataset
ds = load_dataset("tiny-aya-translate/hinglish-casual", split="train", streaming=True)

Streaming is recommended at this size. Licence: upstream terms apply and are not restated here — verify before redistribution.

Code

repo	what it does
`sound-quality-check`	4-stage speech-dataset quality control

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).