Document generation pipeline, speaker diversity, timbre matching, and QA in the dataset card

README.md

@@ -28,3 +28,69 @@ dataset_info:
   download_size: 564580628308
   dataset_size: 582098414656.2
 ---
+
+# PhoMT EN–VI Parallel Speech
+
+Synthetic parallel speech for English↔Vietnamese speech-to-speech translation research:
+**696,243 utterance pairs (~1,200 hours of Vietnamese speech + comparable English)**, generated
+from the sentence-aligned text pairs of [PhoMT](https://huggingface.co/datasets/ura-hcmut/PhoMT).
+Each row carries the English and Vietnamese text plus a spoken rendition of each side.
+
+| | |
+|---|---|
+| Rows | 696,243 |
+| Vietnamese audio | 48 kHz mono WAV, ~1,228 h total (mean ≈ 6.4 s/clip) |
+| English audio | 24 kHz mono WAV |
+| Length balance | EN/VI duration ratio kept within 0.4–1.8 |
+
+```python
+from datasets import load_dataset
+ds = load_dataset("anquachdev/PhoMT-en-vi-speech", split="train", streaming=True)
+row = next(iter(ds))  # en, vi, audio_en, audio_vi, duration_en_s, duration_vi_s, duration_ratio_en_vi
+```
+
+## How it was generated
+
+- **Vietnamese**: [VieNeu-TTS v3-Turbo](https://huggingface.co/pnnbao-ump/VieNeu-TTS) at 48 kHz.
+- **English**: [Kokoro-82M](https://huggingface.co/hexgrad/Kokoro-82M) at 24 kHz.
+- Voices, speaking speed, and target gender are assigned per row by a seeded deterministic RNG,
+  so the corpus is exactly reproducible from the PhoMT indexes.
+
+### Speaker diversity via voice cloning
+
+The Vietnamese voice pool contains **40 voices**: 12 VieNeu presets plus **28 voices cloned from
+[VIVOS](https://huggingface.co/datasets/AILAB-VNUHCM/vivos) speakers** using VieNeu's reference-audio
+voice cloning. Every candidate voice passed a QA scorecard before entering the pool —
+transcription accuracy (PhoWhisper-small CER on synthesized calibration text), clone fidelity
+(speaker-embedding similarity to the reference), and speaking rate; 8 candidates that failed
+(unstable CER or dragging pace) were excluded.
+
+### Cross-lingual timbre matching
+
+To keep speaker identity consistent across the language pair, each row's English voice is derived
+from its Vietnamese voice: both sides are embedded with VieNeu's 192-d speaker encoder (from
+synthesized calibration audio, not raw references), and the Vietnamese voice is matched to its
+nearest same-gender English candidate by cosine similarity. The English candidate grid holds 34
+timbres — 7 Kokoro voices plus pairwise same-gender blends at 25/50/75 mixing weights — of which
+~14 are selected by the matching. Rows generated after this upgrade (~51% of the corpus, PhoMT
+indexes ≥ 345600) use matched pairs; earlier rows pair voices independently.
+
+## Quality control
+
+Every clip was validated before upload:
+
+- **Silence/corruption gate**: each waveform decode-checked — non-finite or near-silent
+  (RMS < 1e-4) output is regenerated (this catches rare GPU-decoder faults).
+- **Length-balance filter**: pairs outside the 0.4–1.8 EN/VI duration ratio band were
+  regenerated once with fresh sampling; pairs still out of band (genuinely imbalanced
+  translations) were excluded.
+- Upload integrity is tracked per source index in `upload-state.json` — every row maps to a
+  unique PhoMT pair, with no duplicates.
+
+## Attribution
+
+Text pairs from [PhoMT](https://github.com/VinAIResearch/PhoMT) (VinAI Research). Vietnamese
+voice references from [VIVOS](https://huggingface.co/datasets/AILAB-VNUHCM/vivos) (AILAB, VNU-HCM).
+Speech synthesized with [VieNeu-TTS](https://huggingface.co/pnnbao-ump/VieNeu-TTS) and
+[Kokoro-82M](https://huggingface.co/hexgrad/Kokoro-82M). The audio is fully synthetic; no human
+recordings are distributed in this dataset.