AkCodes23/sarvam-tts-in-te-en

Indian English + Telugu Single-Speaker TTS Dataset (emotion-tagged)

Clean audio clips sourced from YouTube, transcribed with Sarvam ASR, segmented with diarization, and labeled with emotion/style tags. Built as a data-quality / curation exercise.

"Single-speaker" means each clip contains exactly one speaker (verified by diarization and speaker-embedding similarity). The dataset spans 9 distinct speakers total (4 English, 5 Telugu), tracked via speaker_id.

Contents

• 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: 60.25 minutes.

Evaluation

• 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 candidate speakers flagged).
• Transcript reliability: an independent recogniser (Whisper) agrees with the English Sarvam transcripts at 5.5% word error (n=40). For Telugu, a Telugu-specialised Indic recogniser brings word error against the Sarvam transcripts to 47%, down from 76% with a generic multilingual model; Telugu is best judged by forced alignment and the human listening pass rather than by cross-ASR alone.
• 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 3.09 (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.
• Emotion-label agreement (Krippendorff alpha): 0.44 between the two LLM raters. Adding two acoustic SER models drops the three-rater alpha near zero, because those models cluster toward neutral and were not trained for Telugu, so the emotion labels are treated as weakly supervised and ship with a confidence. Per-clip VAD (valence, arousal, dominance) is included.
• Validation by a larger model (sarvam-105b re-judging the 30b tags, 310 published clips): 89% of transcripts judged clean, 81% suitable to train on, and the emotion endorsed on 26%. Each clip also carries a topic; the set is mostly storytelling (mythology, folk tales, audiobook fiction).
• Audio bandwidth: the source recordings are band-limited (median 99% energy roll-off ~4.1 kHz for English, ~2.6 kHz for Telugu), so although the files are stored at 24 kHz they suit standard 16–24 kHz TTS rather than full-band synthesis.
• Human listening audit (80 clips, 40 per language): 37/40 English and 35/40 Telugu transcripts matched the audio exactly, the rest off by minor punctuation or a single word, and no clip was judged unusable for TTS.

See the project report (GitHub repo) for full methodology and figures.

Schema

audio (24 kHz mono), text (raw transcript), annotated_text (English code-switch spans bracketed, truncation marked with an em dash), normalized_text (language-aware: numbers and abbreviations expanded for TTS reading), language, emotion (neutral, happy, sad, angry, excited, calm, fearful, surprised), style (narrative, conversational, formal, expressive, whisper), emotion_confidence, tag_source (auto/human), topic, speaker_id, gender, accent, duration; quality scores (snr_db, dnsmos_ovrl/sig/bak, dnsmos_pass, squim_*, mms_align_score, overlap_flag, llm_tts_suitable); VAD (valence/arousal/dominance); annotation flags (below); and provenance (source_video_id/url/channel, license, segment_start/end, sample_rate).

Annotation flags

Each clip records what is imperfect about it, so users can filter rather than trust blindly. The two audio-quality flags are automatically inferred proxies, not verified audible-noise labels: quality_flag (a quality concern: DNSMOS < 3.0, or SNR < 18 dB, or elevated energy in pauses) and low_quality_audio (clearly degraded: DNSMOS < 2.8). The rest: has_truncation (ends mid-utterance), has_codemix (preserved English in a regional clip; 0 in practice, since Sarvam ASR transliterates English into Telugu script), has_laughter (audible laughter, set by a listening pass), emotion_low_confidence (tag confidence < 0.55), transcript_review_needed (judge-flagged or alignment < 0.85), overlap_flag (possible second voice). annotation_flags is the pipe-joined list per clip.

Filtering recommendations

• Studio-like subset: dnsmos_pass == true and quality_flag == false and has_truncation == false
• Expressive subset: emotion_confidence > 0.7 and emotion != "neutral"
• Storytelling subset: topic in ('mythology', 'folktale', 'fiction')
• Review queue: transcript_review_needed == true or emotion_low_confidence == true

How it was built

1. Curated single-speaker YouTube sources (audiobooks, lectures, news, storytelling).
2. Sarvam batch STT (saaras:v3) with diarization + timestamps for structure.
3. Silence-snapped segmentation into 3–25 s clips (single speaker only).
4. Sarvam realtime STT (saarika:v2.5) per clip for clip-accurate transcripts.
5. Automated quality gates (clipping, SNR, silence, music/noise bed, ASR confidence, dedup).
6. Hybrid emotion tagging: per-speaker-normalized acoustic features + Sarvam LLM, with an acoustic whisper override.
7. Human review (listen, fix transcripts, relabel); human labels override automated ones.
8. Light loudness normalization (dynamics preserved), balanced emotion selection.

Audio

24 kHz mono WAV. Loudness lightly normalized (~-20 LUFS, peak −1 dBFS) WITHOUT limiting, so the prosodic dynamics that carry emotion are preserved.

Ethics & licensing

Sourced from YouTube for research; clips are short and transformative. Per-clip provenance (source_url, source_channel, license) is retained. Respect the original creators' rights; remove clips on request.

Limitations

Emotion tags are weakly supervised (acoustic features + LLM, partly human-verified) and can miss subtle prosody. The source audio is band-limited, so it suits standard 16–24 kHz TTS rather than full-band synthesis, and the English half leans heavily on one narrator. See the project report for the iteration history.