MeghanaKap/scraped_datasets

Scraped Bengali TTS Speech Chunks

Single-speaker Bengali (bn) speech chunks extracted from YouTube videos, cleaned and segmented for TTS (text-to-speech) training data.

Pipeline

Each chunk was produced by:

1. Downloading source audio via yt-dlp
2. Trimming the first/last 2 minutes of each source video
3. Removing background music/noise via Demucs vocal separation
4. Splitting stereo channels into independent left/right tracks when present
5. Voice Activity Detection (Silero VAD) to find speech regions per track
6. Speaker diarization (pyannote.audio), run once on the full track; only speaker turns with ZERO overlap against any other speaker are kept, then intersected with the VAD speech regions -- chunk boundaries come from these clean single-speaker windows, not from raw VAD segments, so overlapping/crosstalk audio is excluded rather than merely down-weighted
7. Emotion tagging (Hatman/audio-emotion-detection) on each accepted chunk -- metadata only, nothing is filtered by emotion

Columns

• audio: the speech chunk, mono, 16kHz WAV
• language: ISO 639-1 language code (bn = Bengali)
• seconds: duration of the chunk in seconds
• emotion: predicted emotion label (Angry, Disgusted, Fearful, Happy, Neutral, Sad, Suprised) from Hatman/audio-emotion-detection

Config

• bengali: all Bengali-language chunks collected so far