Add dataset card

README.md

@@ -4,77 +4,13 @@ language:
 license: cc-by-4.0
 task_categories:
 - automatic-speech-recognition
-pretty_name: FLEURS — Mongolian (Clean)
+task_ids: []
+pretty_name: "FLEURS — Mongolian (Clean)"
 tags:
 - mongolian
 - speech
 - audio
 - fleurs
-configs:
-- config_name: default
-  data_files:
-  - split: train
-    path: data/train-*
-  - split: validation
-    path: data/validation-*
-  - split: test
-    path: data/test-*
-dataset_info:
-  features:
-  - name: audio
-    dtype:
-      audio:
-        sampling_rate: 24000
-  - name: snr_db
-    dtype: float32
-  - name: mean_f0_hz
-    dtype: float32
-  - name: pitch_confidence
-    dtype: float32
-  - name: dnsmos_sig
-    dtype: float32
-  - name: dnsmos_bak
-    dtype: float32
-  - name: dnsmos_ovr
-    dtype: float32
-  - name: dnsmos_p808
-    dtype: float32
-  - name: cer
-    dtype: float32
-  - name: asr_transcript
-    dtype: string
-  - name: duration_s
-    dtype: float32
-  - name: id
-    dtype: int32
-  - name: num_samples
-    dtype: int32
-  - name: path
-    dtype: string
-  - name: raw_transcription
-    dtype: string
-  - name: transcription
-    dtype: string
-  - name: gender
-    dtype: int32
-  - name: lang_id
-    dtype: int32
-  - name: language
-    dtype: string
-  - name: lang_group_id
-    dtype: int32
-  splits:
-  - name: train
-    num_bytes: 1005988185
-    num_examples: 2541
-  - name: validation
-    num_bytes: 122753848
-    num_examples: 357
-  - name: test
-    num_bytes: 314641404
-    num_examples: 854
-  download_size: 1441625360
-  dataset_size: 1443383437
 ---
 
 # FLEURS — Mongolian (Quality-Filtered)
@@ -94,12 +30,12 @@ Derived from `google/fleurs` config `mn_mn`. FLEURS is the speech version of the
 | 1. Format normalization | librosa | mono · 16 kHz |
 | 2. Voice activity detection | Silero VAD | ≥25 % speech frames |
 | 3. SNR filter | RMS-based SNR | ≥8 dB |
-| 4. Pitch & clarity | CREPE F0 | F0 ≥50 Hz · confidence ≥0.25 |
+| 4. Pitch metadata | CREPE F0 | recorded when available; not a rejection gate |
 | 5. AI quality score | DNSMOS P.835 | OVR ≥2.2 · SIG ≥2.4 · BAK ≥2.0 |
-| 6. Full sentence verification | Whisper large-v3 + CER | CER ≤0.35 · length ratio ≥0.40 |
+| 6. Full sentence verification | Whisper large-v3 + CER | CER ≤0.35, or ≤0.50 when length ratio is 0.75–1.25 |
 
 Ground truth for sentence verification: `raw_transcription` field.
-All passing clips peak-normalized to −1 dBFS and resampled to **24 kHz**.
+Clips are kept between **1–30 seconds** to match oron-tts training limits. All passing clips are peak-normalized to −1 dBFS and resampled to **24 kHz**.
 
 ## Schema
 
@@ -110,7 +46,7 @@ All original FLEURS fields preserved, plus computed quality metrics:
 | `id` | int32 | Sample ID |
 | `num_samples` | int32 | Number of audio samples |
 | `path` | string | Audio file path |
-| `audio` | Audio(16000) | Decoded audio at 16 kHz |
+| `audio` | Audio(24000) | Cleaned audio resampled to 24 kHz |
 | `raw_transcription` | string | Original (unnormalized) transcription |
 | `transcription` | string | Normalized transcription |
 | `gender` | int32 | Speaker gender class |
@@ -132,7 +68,7 @@ All original FLEURS fields preserved, plus computed quality metrics:
 
 ```python
 from datasets import load_dataset
-ds = load_dataset("btsee/fleurs-mn", "mn_mn")
+ds = load_dataset("btsee/fleurs-mn")
 sample = ds["train"][0]
 ```