Upload README.md with huggingface_hub

README.md

@@ -33,183 +33,6 @@ configs:
     path: telugu/validation-*
   - split: test
     path: telugu/test-*
-dataset_info:
-- config_name: indian_english
-  features:
-  - name: audio
-    dtype:
-      audio:
-        sampling_rate: 24000
-  - name: text
-    dtype: string
-  - name: normalized_text
-    dtype: string
-  - name: language
-    dtype: string
-  - name: language_code
-    dtype: string
-  - name: emotion
-    dtype: string
-  - name: style
-    dtype: string
-  - name: emotion_confidence
-    dtype: float32
-  - name: tag_source
-    dtype: string
-  - name: speaker_id
-    dtype: string
-  - name: gender
-    dtype: string
-  - name: accent
-    dtype: string
-  - name: duration
-    dtype: float32
-  - name: snr_db
-    dtype: float32
-  - name: dnsmos_ovrl
-    dtype: float32
-  - name: dnsmos_sig
-    dtype: float32
-  - name: dnsmos_bak
-    dtype: float32
-  - name: dnsmos_pass
-    dtype: bool
-  - name: squim_stoi
-    dtype: float32
-  - name: squim_pesq
-    dtype: float32
-  - name: squim_sisdr
-    dtype: float32
-  - name: mms_align_score
-    dtype: float32
-  - name: overlap_flag
-    dtype: bool
-  - name: ser_emotion
-    dtype: string
-  - name: valence
-    dtype: float32
-  - name: arousal
-    dtype: float32
-  - name: dominance
-    dtype: float32
-  - name: topic
-    dtype: string
-  - name: llm_tts_suitable
-    dtype: float32
-  - name: source_video_id
-    dtype: string
-  - name: source_url
-    dtype: string
-  - name: source_channel
-    dtype: string
-  - name: license
-    dtype: string
-  - name: segment_start
-    dtype: float32
-  - name: segment_end
-    dtype: float32
-  - name: sample_rate
-    dtype: int32
-  splits:
-  - name: train
-    num_bytes: 78768129.0
-    num_examples: 144
-  - name: validation
-    num_bytes: 4354531.0
-    num_examples: 8
-  - name: test
-    num_bytes: 3860255.0
-    num_examples: 8
-  download_size: 86088473
-  dataset_size: 86982915.0
-- config_name: telugu
-  features:
-  - name: audio
-    dtype:
-      audio:
-        sampling_rate: 24000
-  - name: text
-    dtype: string
-  - name: normalized_text
-    dtype: string
-  - name: language
-    dtype: string
-  - name: language_code
-    dtype: string
-  - name: emotion
-    dtype: string
-  - name: style
-    dtype: string
-  - name: emotion_confidence
-    dtype: float32
-  - name: tag_source
-    dtype: string
-  - name: speaker_id
-    dtype: string
-  - name: gender
-    dtype: string
-  - name: accent
-    dtype: string
-  - name: duration
-    dtype: float32
-  - name: snr_db
-    dtype: float32
-  - name: dnsmos_ovrl
-    dtype: float32
-  - name: dnsmos_sig
-    dtype: float32
-  - name: dnsmos_bak
-    dtype: float32
-  - name: dnsmos_pass
-    dtype: bool
-  - name: squim_stoi
-    dtype: float32
-  - name: squim_pesq
-    dtype: float32
-  - name: squim_sisdr
-    dtype: float32
-  - name: mms_align_score
-    dtype: float32
-  - name: overlap_flag
-    dtype: bool
-  - name: ser_emotion
-    dtype: string
-  - name: valence
-    dtype: float32
-  - name: arousal
-    dtype: float32
-  - name: dominance
-    dtype: float32
-  - name: topic
-    dtype: string
-  - name: llm_tts_suitable
-    dtype: float32
-  - name: source_video_id
-    dtype: string
-  - name: source_url
-    dtype: string
-  - name: source_channel
-    dtype: string
-  - name: license
-    dtype: string
-  - name: segment_start
-    dtype: float32
-  - name: segment_end
-    dtype: float32
-  - name: sample_rate
-    dtype: int32
-  splits:
-  - name: train
-    num_bytes: 76476710.0
-    num_examples: 134
-  - name: validation
-    num_bytes: 5370070.0
-    num_examples: 8
-  - name: test
-    num_bytes: 4991104.0
-    num_examples: 8
-  download_size: 85878414
-  dataset_size: 86837884.0
 ---
 
 # Indian English + Telugu Single-Speaker TTS Dataset (emotion-tagged)
@@ -222,19 +45,19 @@ diarization, and labeled with emotion/style tags. Built as a data-quality / cura
 > total** (5 English, 6 Telugu), tracked via `speaker_id`.
 
 ## Contents
-- **Indian English** (`indian_english`): 30.01 min, 169 clips, 4 speakers; emotions: {'angry': 31, 'neutral': 31, 'calm': 31, 'sad': 31, 'excited': 31, 'fearful': 8, 'happy': 6}
-- **Telugu** (`telugu`): 30.05 min, 156 clips, 5 speakers; emotions: {'calm': 29, 'neutral': 28, 'angry': 28, 'excited': 28, 'sad': 28, 'happy': 6, 'fearful': 6, 'surprised': 3}
+- **Indian English** (`indian_english`): 30.17 min, 160 clips, 4 speakers; emotions: {'angry': 30, 'neutral': 29, 'calm': 29, 'sad': 29, 'excited': 29, 'fearful': 8, 'happy': 6}
+- **Telugu** (`telugu`): 30.08 min, 150 clips, 5 speakers; emotions: {'calm': 27, 'neutral': 27, 'angry': 27, 'excited': 27, 'sad': 27, 'fearful': 6, 'happy': 6, 'surprised': 3}
 
-Total: **60.06 minutes**.
+Total: **60.25 minutes**.
 
 ## Evaluation (evidence, not just claims)
 
 - **Single-speaker check** (ECAPA-TDNN embeddings): same-speaker cosine 0.74 vs different-speaker 0.21 (separation 0.52, verification AUC 0.96 / EER 9.1%; 0/11 speakers flagged).
-- **Transcript reliability**: English cross-ASR agreement with Whisper = 6.8% WER / 4.5% CER (n=40), strong. Realtime ASR language-ID matched the target language on 100% of EN and 100% of TE clips. Telugu cross-ASR is not a valid proxy (Whisper is weak in Telugu); Telugu transcripts are best audited by human review.
+- **Transcript reliability**: English cross-ASR agreement with Whisper = 5.5% WER / 3.4% CER (n=40), strong. Telugu cross-ASR is not a valid proxy (Whisper is weak in Telugu); Telugu transcripts are best audited by human review.
 - **Emotion-tag reliability** (sarvam-30b vs sarvam-105b on 120 clips): 65% agreement, Cohen's κ 0.55.
 - **Phoneme coverage**: English 39 (100%), Telugu 44 (88%).
-- **Perceptual quality** (DNSMOS OVRL, published set): EN 3.21 (86% pass>3.0), TE 3.19 (86% pass>3.0). Filter `dnsmos_pass=True` for a stricter subset.
-- **Transcript–audio alignment** (MMS forced-align): median confidence EN 0.948, TE 0.934.
+- **Perceptual quality** (DNSMOS OVRL, published set): EN 3.09 (58% pass>3.0), TE 3.16 (81% pass>3.0). Filter `dnsmos_pass=True` for a stricter subset.
+- **Transcript–audio alignment** (MMS forced-align): median confidence EN 0.954, TE 0.937.
 
 See the project report (GitHub repo) for full methodology and figures.