Praxel/codeswitch-pairs-lase-indian

Codeswitch Pairs LASE — Indian-accent held-out corpus

1369 held-out cross-script utterance pairs from 8 ElevenLabs Indian-English Multilingual voices. Surfaces the accent-conditional finding: off-the-shelf encoders cluster Indian-accent voices closely regardless of script, while Western voices show large script-conditional gaps.

Each row is one synthesized utterance with metadata; pairs are reconstructed at evaluation time by joining on voice_id (same voice, different script = cross-script pair).

Schema (manifest.jsonl)

{
  "voice_id": "21m00Tcm4TlvDq8ikWAM",
  "lang": "en | hi | te | ta",
  "text": "the prompt text",
  "wav_path": "<voice_name>/<lang>_NNN.wav",
  "quality": {"cosine": 0.97, "utmos": 4.6}
}

Audio is 16 kHz mono WAV, ~2 s/utterance.

Stats

• Pairs: 1369
• Voices: 8
• Languages: en, hi, te, ta
• Quality gate: WavLM-cosine ≥ 0.90 vs the voice's reference clip

Provenance

Synthesized using the ElevenLabs Multilingual v2 API. Voice IDs are public ElevenLabs catalog voices used under their TOS for research/evaluation purposes. Text prompts are short generic English phrases translated/transliterated into each target script.

Companion paper

LASE: Language-Adversarial Speaker Encoding for Indic Cross-Script Identity Preservation (arXiv:TBD)

License

CC-BY-4.0.