oscowlai/synthetic_S2S

Synthetic S2S Dataset

Languages covered (10): English, Spanish, French, Dutch, Japanese, Arabic, Chinese, Hindi, Telugu, Tamil

This dataset supports training for ASR, MT, TTS, and speech-to-speech translation across the 10 languages above. No public corpus has real same-speaker recordings across all of these language pairs, so most of this dataset is generated: real Hindi audio is the only fully natural side, and every other language's audio is produced via voice-cloning TTS conditioned on that same real Hindi clip — so the same speaker identity carries across all 10 languages for a given row. What follows explains exactly how that mapping works, how the data is organized, and what each field means.

How the mapping works

1. We start from one real dataset — ai4bharat/SeamlessAlign (Indic SeamlessAlign), Hindi split — giving real Hindi audio, real Hindi text, and real English text (already translated in the source).
2. For each Hindi row, we generate 8 more target texts (es, fr, nl, ja, ar, zh, te, ta) and 8 more target audios (voice-cloned TTS, always conditioned on the original real Hindi audio, never on another synthetic clip, to avoid quality loss from cloning a clone).
3. This gives every row 9 parallel (text, audio) tuples in total — 1 real (Hindi) + 8 synthetic (English's text came free from the source, its audio still generated).
4. Any 2 of those 9 tuples can be combined into a training pair — e.g. es→ja is just (text_es, audio_es) as source and (text_ja, audio_ja) as target, assembled at read time. No new translation or TTS call is needed for these cross pairs.
5. From 9 generations per row, this produces all 45 language pairs (90 directed pairs) — the combination step is free; only the per-row generation step costs API/compute time.

All synthetic audio for a row shares the same underlying Hindi speaker's cloned voice and the same gender_label — computed once per Hindi row, not recomputed per pair.

Repository structure

oscowlai/synthetic_S2S/
└── indic_seamlessalign/
    ├── hi-en/
    ├── hi-es/
    ├── hi-fr/
    ├── hi-nl/
    ├── hi-ja/
    ├── hi-ar/
    ├── hi-zh/
    ├── hi-te/
    ├── hi-ta/
    └── [the other 36 combined pairs are assembled at read-time from the above — not stored separately]

Each hi-X/ folder contains parquet shards (generated-00000.parquet, generated-00001.parquet, …), 500 rows each.

Data fields

Field	Type	Description
row_id	int	Shared join key across all hi-X/ folders — same row_id in hi-es and hi-fr always traces back to the same original Hindi sentence. Kept deliberately minimal (just an integer) so cross-pair combination stays reliable even if some rows were skipped in one pair but not another.
source_lang	string	Source language code for this row
target_lang	string	Target language code for this row
source_audio	bytes (wav)	Source-side audio
target_audio	bytes (wav)	Target-side audio (voice-cloned TTS for all pairs except real recordings)
source_text	string	Source-side transcript
target_text	string	Target-side translated text
gender_label	string	male / female / unknown — pitch-based heuristic (librosa YIN), computed once per Hindi row and reused across all pairs derived from it
reference_audio	bytes (wav)	The original real Hindi clip used to condition voice-cloning. Identical to source_audio for hi-X rows; distinct from source_audio for combined non-Hindi pairs (e.g. in es-ja, reference_audio is still the original Hindi clip, not the Spanish source_audio)

Translation method per pair

Target language	Method
en	Native to source dataset — no translation needed
te, ta	IndicTrans2 (ai4bharat/indictrans2-indic-indic-dist-320M), run locally
es, fr, nl, ja, ar, zh	Google Translate

Mapping

Step 1 — Generate from Hindi only. Every Hindi sentence is translated and voice-cloned into the other 9 languages. This is the only step that costs translation/TTS work — done once per sentence, 9 times.

From	To	Work done
Hindi (real)	English	Translate + clone voice
Hindi (real)	Spanish	Translate + clone voice
Hindi (real)	French	Translate + clone voice
Hindi (real)	Dutch	Translate + clone voice
Hindi (real)	Japanese	Translate + clone voice
Hindi (real)	Arabic	Translate + clone voice
Hindi (real)	Chinese	Translate + clone voice
Hindi (real)	Telugu	Translate + clone voice
Hindi (real)	Tamil	Translate + clone voice

Step 2 — Every pair among all 10 languages is just picking 2 from the list above. No further generation. This covers Hindi paired with any of the 9, and all 9 non-Hindi languages paired with each other, in both directions.

	EN	ES	FR	NL	JA	AR	ZH	TE	TA
HI	✓	✓	✓	✓	✓	✓	✓	✓	✓
EN	—	✓	✓	✓	✓	✓	✓	✓	✓
ES		—	✓	✓	✓	✓	✓	✓	✓
FR			—	✓	✓	✓	✓	✓	✓
NL				—	✓	✓	✓	✓	✓
JA					—	✓	✓	✓	✓
AR						—	✓	✓	✓
ZH							—	✓	✓
TE								—	✓
TA									—

Every ✓ = a usable pair, in both directions, assembled at read-time from Step 1's output — nothing in this grid required a new translation or TTS call.

Source data attribution

Base real data: Indic SeamlessAlign (ai4bharat/SeamlessAlign, indic2en config, hindi split), part of the BhasaAnuvaad project (AI4Bharat). Synthetic audio generated via Fish Audio (s2.1-pro-free).