Update README.md

README.md

@@ -768,14 +768,7 @@ configs:
 
 **53,635 hours of transcribed speech across 113 language–region configs**, drawn from national parliaments, public broadcasters, public-domain audiobooks, and international institutions (UN, ICC, Vatican Radio).
 
-Every segment consists of a human transcript, an aligned ASR transcript, character error rate (CER) between the two, a WADA-SNR estimate, and four DNSMOS-P.835 quality scores.
-
-**Audio format:** OGG Opus, 24 kHz mono, ~64 kbps VBR
-**Splits:** 95 / 5 train / test (per config, seed 42)
-**Provenance fields:** `source`, `source_url`, `session_date`, `source_start_s`, `source_end_s`, `segment_id` on every row
-**License:** CC BY-NC 4.0, with per-config restrictions where the upstream license is more restrictive (see [Per-config licenses](#per-config-licenses))
-
-> ⚠️ **Audio embedding caveat.** Five configs (`kh_km` Khmer, `la_lo` Lao, `mm_my` Burmese, `vn_vi` Vietnamese, `cn_ug` Uyghur 3,941 hours total, all Radio Free Asia–sourced) include only metadata, transcripts, and source URLs. Audio for these configs must be downloaded separately from `source_url`. **Embedded-audio hours: ~49,694.** Total hours including metadata-only configs: 53,635.
+Every row consists of a 24kHz speech utterance paired with a human transcript, an aligned ASR transcript, character error rate (CER) between the two, a WADA-SNR estimate, and four DNSMOS-P.835 quality scores.
 
 ---
 
@@ -852,14 +845,6 @@ ds_tts = ds.filter(lambda row:
 )
 ```
 
-Recommended starting thresholds (tune per use case):
-
-| Use case             | DNSMOS OVR | SNR (dB) | CER    |
-| -------------------- | ---------- | -------- | ------ |
-| TTS / voice cloning  | ≥ 3.2      | ≥ 15     | ≤ 0.05 |
-| Clean ASR            | ≥ 2.8      | ≥ 10     | ≤ 0.10 |
-| Robustness / noisy   | —          | —        | ≤ 0.15 |
-
 ---
 
 ## Schema
@@ -887,21 +872,6 @@ Recommended starting thresholds (tune per use case):
 
 ---
 
-## Why WorldSpeech
-
-| Corpus           | Hours        | Languages | Human transcripts | Domain                     |
-| ---------------- | ------------ | --------- | ----------------- | -------------------------- |
-| MLS              | ~50k         | 8         | ✓                 | Audiobooks                 |
-| VoxPopuli        | ~24k (transcribed) | 23  | Partial           | Parliamentary (EU only)    |
-| GigaSpeech 2     | ~30k         | 3         | ✗                 | YouTube                    |
-| Common Voice 17  | ~31k         | 100+      | ✓                 | Read prompts               |
-| FLEURS           | ~1.4k        | 102       | ✓                 | Read (eval only)           |
-| **WorldSpeech**  | **~53,635** | **79** | **✓ (+ ASR + CER)** | **Parliamentary, broadcast, audiobook** |
-
-WorldSpeech combines MLS-scale hours with FLEURS-scale language coverage and per-row quality metadata that neither provides, with a deliberate focus on languages outside the EU.
-
----
-
 ## Recommended use cases
 
 - **Multilingual ASR pretraining and fine-tuning**, especially for languages with no MLS / VoxPopuli coverage.
@@ -1091,3 +1061,6 @@ Issues, takedowns, license questions: open a discussion on this repository.
 ## Per-config licenses
 
 *(See dataset card for full per-config license table.)*
+
+> ⚠️ **Audio embedding caveat.** Five configs (`kh_km` Khmer, `la_lo` Lao, `mm_my` Burmese, `vn_vi` Vietnamese, `cn_ug` Uyghur 3,941 hours total, all Radio Free Asia–sourced) include only metadata, transcripts, and source URLs. Audio for these configs must be downloaded separately from `source_url`. **Embedded-audio hours: ~49,694.** Total hours including metadata-only configs: 53,635.
+