README: samelang_expressive (in-progress, ~3.3 TB) + samples_showcase as landing preview

README.md

@@ -42,7 +42,7 @@ multi-lingual voice-clone models. Covers 4 target languages
 via sglang-omni.
 
 > **🚧 Upload in progress** — the main `samelang_expressive` corpus
-> (~4.78 M rows / ~1.4 TB) is being pushed as parquet shards.
+> (~4.78 M rows / **~3.3 TB**) is being pushed as parquet shards.
 > Expect the shard count to grow over the next hours/days until each
 > language reaches `train-XXXXX-of-00240.parquet`.
 
@@ -67,16 +67,16 @@ synthesize **1000 target-language texts**, giving
 `1200 speakers × 4 langs × 1000 texts ≈ 4.8 M` cloned wavs.
 
 - **Shards**: `data/samelang/<lang>/train-XXXXX-of-00240.parquet`
-  (each lang → 240 shards, ~5000 rows/shard, ~1.5 GB with zstd)
+  (each lang → 240 shards, ~5000 rows/shard)
 - **Row schema**: `audio` (struct{bytes, path}, auto-cast to HF `Audio`
   feature) + `text`, `lang`, `ref_id`, `speaker_id`, `text_id`,
   `duration`, `sample_rate`, `ref_audio_path`, `ref_text`, `engine`,
   `original_text`
-- **Total size (fully uploaded)**: ~1.4 TB audio, ~4.78 M rows, ~11 k h
+- **Total size (fully uploaded)**: **~3.3 TB** audio, ~4.78 M rows, ~11 k h
 - **Engine**: `sglang-omni-higgs-audio-v3` with expressive relabel-based
   prompt (drawn from the top-quality pool per source speaker)
 
-### Streaming (recommended for 1.4 TB)
+### Streaming (recommended for 3.3 TB)
 
 ```python
 from datasets import load_dataset
@@ -123,124 +123,87 @@ hf download leeoxiang/multi_lingo_data \
 
 # Multi-lingual TTS Data (leeoxiang/multi_lingo_data)
 
-Splits:
-
-1. **`references`** — real Emilia mono-lingual speaker clips (source of
-   truth for voice cloning).
-2. **audio_prompts variants** — synthesized cross-lingual clips: each
-   real speaker cloned to speak every target language. Used as
-   pseudo-refs for Step 2 same-lang expansion.
-   - **`audio_prompts`** — 14400 wavs
-   - **`audio_prompts_expressive`** — 14400 wavs
-
-## References (real Emilia speakers)
-
-Curated real-speaker reference clips selected from Emilia by
-``scripts/select_references.py``. Each row of ``references.jsonl``
-points to its wav via ``audio_file`` (relative to repo root).
-
-- Total refs: **1200**
-- Total duration: **2.85 h**
-
-| lang | n_refs |
-|------|--------|
-| en | 300 |
-| ja | 300 |
-| ko | 300 |
-| zh | 300 |
-
-
-## Audio Prompts — `audio_prompts`
-
-Generated by ``scripts/generate_audio_prompts_higgs_sglang.py`` against
-a local sglang-omni server serving ``bosonai/higgs-tts-3-4b``. For each
-reference speaker × target language × text prompt, one wav.
-
-- Total wavs: **14400**
-- Total duration: **59.56 h**
-- Engine: `sglang-omni-higgs-audio-v3`
-- Layout: `audio_prompts/<target_lang>/<ref_id>/<text_id:04d>.wav`
-
-### Per target-language
-
-| target_lang | n_wavs |
-|-------------|--------|
-| en | 3600 |
-| ja | 3600 |
-| ko | 3600 |
-| zh | 3600 |
-
-### Per source-language (Emilia speaker's native lang)
-
-| source_lang | n_wavs |
-|-------------|--------|
-| en | 3600 |
-| ja | 3600 |
-| ko | 3600 |
-| zh | 3600 |
-
-
-## Audio Prompts — `audio_prompts_expressive`
+Large-scale cross-lingual + same-lingual TTS corpus for training expressive
+multi-lingual voice-clone models. Covers 4 target languages
+(**en / ja / ko / zh**), synthesized by `bosonai/higgs-tts-3-4b`
+via sglang-omni.
 
-Generated by ``scripts/generate_audio_prompts_higgs_sglang.py`` against
-a local sglang-omni server serving ``bosonai/higgs-tts-3-4b``. For each
-reference speaker × target language × text prompt, one wav.
+> **🚧 Upload in progress** — the main `samelang_expressive` corpus
+> (~4.78 M rows / ~1.4 TB) is being pushed as parquet shards.
+> Expect the shard count to grow over the next hours/days until each
+> language reaches `train-XXXXX-of-00240.parquet`.
 
-- Total wavs: **14400**
-- Total duration: **59.84 h**
-- Engine: `sglang-omni-higgs-audio-v3`
-- Layout: `audio_prompts_expressive/<target_lang>/<ref_id>/<text_id:04d>.wav`
+## 🎧 Listen first — `samples_showcase`
 
-### Per target-language
+40 randomly-picked cross-lingual sample pairs (10 per target language)
+from the `samelang_expressive` corpus, each paired with the *reference
+wav* used for voice cloning so you can A/B the timbre. This is what the
+HF viewer above shows by default.
 
-| target_lang | n_wavs |
-|-------------|--------|
-| en | 3600 |
-| ja | 3600 |
-| ko | 3600 |
-| zh | 3600 |
+```python
+from datasets import load_dataset
+ds = load_dataset("leeoxiang/multi_lingo_data", "samples_showcase", split="train")
+print(ds[0]["text"], ds[0]["audio"]["sampling_rate"])
+```
 
-### Per source-language (Emilia speaker's native lang)
+## 📦 Main corpus — `samelang_expressive` (**~4.78 M rows, in progress**)
 
-| source_lang | n_wavs |
-|-------------|--------|
-| en | 3600 |
-| ja | 3600 |
-| ko | 3600 |
-| zh | 3600 |
+Same-language expansion: for every pseudo-reference (real Emilia speaker
+cross-cloned into the target language via `audio_prompts_expressive`), we
+synthesize **1000 target-language texts**, giving
+`1200 speakers × 4 langs × 1000 texts ≈ 4.8 M` cloned wavs.
 
+- **Shards**: `data/samelang/<lang>/train-XXXXX-of-00240.parquet`
+  (each lang → 240 shards, ~5000 rows/shard, ~1.5 GB with zstd)
+- **Row schema**: `audio` (struct{bytes, path}, auto-cast to HF `Audio`
+  feature) + `text`, `lang`, `ref_id`, `speaker_id`, `text_id`,
+  `duration`, `sample_rate`, `ref_audio_path`, `ref_text`, `engine`,
+  `original_text`
+- **Total size (fully uploaded)**: ~1.4 TB audio, ~4.78 M rows, ~11 k h
+- **Engine**: `sglang-omni-higgs-audio-v3` with expressive relabel-based
+  prompt (drawn from the top-quality pool per source speaker)
 
-### CSV schema (per-variant ``<subdir>/metadata.csv``)
+### Streaming (recommended for 1.4 TB)
 
-| column           | description                                       |
-|------------------|---------------------------------------------------|
-| file_name        | wav path relative to <subdir>/metadata.csv        |
-| ref_id           | source speaker key                                |
-| ref_lang         | source speaker's native language                  |
-| speaker_id       | original Emilia speaker id                        |
-| lang             | target language of this synth                     |
-| text_id          | id in ``texts/<lang>_text.jsonl``                 |
-| text             | target-language prompt text                       |
-| duration         | seconds                                           |
-| sample_rate      | Hz (higgs-audio-v3 outputs 24000)                 |
-| ref_audio_path   | absolute path to the reference wav (server-side)  |
-| ref_text         | transcript of the reference clip                  |
-| ref_dnsmos       | DNSMOS score of the reference                     |
-| engine           | fixed to `sglang-omni-higgs-audio-v3`             |
+```python
+from datasets import load_dataset
+ds = load_dataset(
+    "leeoxiang/multi_lingo_data",
+    "samelang_expressive",
+    split="train",
+    streaming=True,
+)
+row = next(iter(ds))
+print(row["lang"], row["text"][:40], row["audio"]["sampling_rate"])
+```
 
-## Loading
+### Single language
 
 ```python
 from datasets import load_dataset
+ds = load_dataset(
+    "leeoxiang/multi_lingo_data",
+    data_files="data/samelang/zh/train-*.parquet",
+    split="train",
+    streaming=True,
+)
+```
 
-# References only
-refs = load_dataset("leeoxiang/multi_lingo_data", "references", split="train")
+### Full download
 
-# Audio prompts (cross-lingual synths)
-audio_prompts = load_dataset("leeoxiang/multi_lingo_data", "audio_prompts", split="train")
-audio_prompts_expressive = load_dataset("leeoxiang/multi_lingo_data", "audio_prompts_expressive", split="train")
+```bash
+hf download leeoxiang/multi_lingo_data \
+    --repo-type dataset \
+    --include 'data/samelang/**' \
+    --local-dir ./multi_lingo_data
 ```
 
-<!-- viewer-refresh -->
+## Supporting configs
+
+- **`references`** — 1200 curated Emilia mono-lingual real-speaker clips.
+- **`audio_prompts`** — 14.4 K cross-lingual synth from Step 1 (each
+  real Emilia speaker cloned into 4 target langs × 3 texts).
+- **`audio_prompts_expressive`** — same as above with expressive prompt
+  selection; the pseudo-refs used to seed `samelang_expressive`.
 
 </details>