💜 Support TeraTTS

If this dataset is useful to you, consider supporting TeraTTS development. Thank you for your support!

Currency	Donation address
ETH	0x19766c207F82303c3998976d852DB82e8Dcc5d77
GRAM (TON)	UQDYx1blVO8H31APnRUXVg2uTXV--wXX8HdTyKxnOKQciPQj
SOL	DvBjAqaj4AYWi3JBMcW4FmXsaU4Nu6yjxrox4XJZ6o6Z
Tether USDt (TRON / TRC20)	TZ6QeYZSr2h22rzGqQZQZTMzWYVPdUej1y
BTC	bc1q2l0xdlylqpweu4sm3y53f95g0nazcm8yu42gkq

---

Webinars

315,085 audio segments with ASR text · 2,064.61 hours · 2,319 source recordings.

Russian educational webinars, filtered for language and detected music, prepared as self-contained Parquet shards for speech recognition and speech synthesis research. Each row contains an audio segment and its automatically recognized text. Audio players are available in the dataset viewer above.

Source audio: ZeroAgency/shkolkovo-bobr.video-webinars-audio, from bobr.video and Школково. All credit for the original recordings and course materials goes to their creators.

Contents

Property	Value
Split	train
Audio/text pairs	315,085
Audio hours	2,064.61
Source recordings represented	2,319
Shards	965 Parquet files
Download size	255.76 GiB
Audio	Mono FLAC, 16-bit PCM, native 44.1 or 48 kHz
ASR language label	ru for all 315,085 segments

Audio bytes are embedded in Parquet; no source MP3 downloads or local paths are needed. The train split includes all published pairs. This release does not define a held-out evaluation set. Split by recording_id when constructing your own validation set.

Load from Hugging Face

from datasets import load_dataset

ds = load_dataset("TeraTTS/webinars", split="train", streaming=True)
sample = next(iter(ds))
print(sample["text"])
print(sample["duration"], sample["sample_rate"])

# Current datasets versions expose a TorchCodec decoder:
audio = sample["audio"].get_all_samples()
waveform = audio.data
sample_rate = audio.sample_rate

For access to the embedded FLAC without an audio decoder:

from datasets import Audio, load_dataset
import io
import soundfile as sf

ds = load_dataset("TeraTTS/webinars", split="train", streaming=True)
ds = ds.cast_column("audio", Audio(decode=False))
sample = next(iter(ds))
waveform, sample_rate = sf.read(io.BytesIO(sample["audio"]["bytes"]))
print(sample["text"])

Columns

Column	Meaning
audio	Embedded FLAC, represented by Hugging Face's Audio feature
text	Original Whisper ASR text for the segment
id	Stable segment identifier
duration	Duration of the published audio in seconds
language	Automatic Whisper language label; may be incorrect
recording_id	Upstream webinar UUID, shared by segments from one recording
sample_rate	Native audio sample rate
start_seconds, end_seconds	Segment's enclosing interval in the original MP3
source_spans	Ordered native sample intervals retained from the original MP3
removed_pause_seconds	Time removed from internal quiet pauses

Each source_spans entry has start_sample and end_sample, with an exclusive end. Decode the original stereo MP3, take the arithmetic channel mean, then concatenate these intervals to reproduce the segment before PCM16 conversion. The published duration can be shorter than end_seconds - start_seconds. The original webinar can be looked up at https://bobr.video/watch/<recording_id>.

Preparation

1. Original audio was downloaded from ZeroAgency/shkolkovo-bobr.video-webinars-audio at revision c35640ea983658baa75047ad0d058c7912ff08ce. This release uses the recordings, with newly generated local transcripts; it does not reuse the source repository's TXT transcripts.
2. FireRed VAD located speech and grouped it into segments targeting 20–30 seconds. Short tails remain; some boundaries were forced to satisfy the maximum duration.
3. ASR used Systran/faster-whisper-large-v3, revision edaa852ec7e145841d8ffdb056a99866b5f0a478, with FP16 inference, beam size 5, automatic language detection and temperature fallback.
4. Long internal quiet pauses were shortened using the saved, audited edit maps. Eligible quiet runs of at least 0.8 s retained 0.3 s, protecting VAD speech with 60 ms margins and trusted Unity2 word alignments with 80 ms margins. Untrusted alignments used a more conservative energy threshold. Edge silence was retained. No new alignment or transcription was run for this publication.
5. The ASR text was preserved verbatim. The 130 segments with empty or whitespace-only transcripts were omitted. The language/music filters below were then applied. No stress marks or prosody tags were added to the transcripts.
6. Audio was reconstructed at its native sample rate, converted to mono PCM16 and encoded as FLAC. Every original and pause-edited float32 chunk was checked against the waveform hashes saved during data preparation. FLAC headers, lengths and representative decode round trips were checked before upload.

FLAC stores the exported PCM16 signal losslessly; the source recordings were already MP3. This export does not restore frequencies or detail lost by MP3. Full shard checksums and measured counts are in manifest.json.

Language and music filtering

This filtered release retains only segments whose existing ASR language tag is ru. It removes 3,526 other-tagged segments and 8,643 additional Russian-tagged segments flagged for music, background music, singing, songs or musical instruments.

EfficientAT mn10_as scanned every Russian-tagged segment using 10-second windows with a 5-second hop, including the final window. Short clips were zero-padded. Analysis used temporary 32 kHz mono waveforms reconstructed from the same saved pause-edit spans as the release. AST independently checked all primary candidates (score >= 0.05 or background-music score >= 0.005) and 1,000 deterministic low-score audit samples at 16 kHz. Where these two detectors disagreed at a maximum target-class/window score of 0.15, EfficientAT mn40_as resolved the decision. A segment was removed when at least two detectors exceeded that threshold. The low-score audit found no additional music decisions under this rule.

All shards were rebuilt from local source recordings and audited ASR/edit records. Retained audio signals, transcripts, identifiers and source spans are unchanged. Detector revisions, thresholds and counts are in filtering.json; per-sample decisions and scores are in filter_scores.parquet. The unfiltered version remains reproducible at revision 83c2459778926a7db6144abd552b084de1ac95e9. Filtering is automatic; quiet music can remain and false exclusions are possible. Detector scores are not calibrated probabilities of a clean recording.

Transcript quality

These are automatic transcripts, without manual correction or a measured corpus WER. Retained recordings can contain multiple speakers, background sounds, filler words, course-specific names or formulas, and music missed by the automatic filter. ASR may hallucinate, omit words or misidentify a language. Speech-aware segmenting and pause checks do not establish transcript correctness. This release preserves those limitations so users can apply task-specific filtering.

Source and license

The upstream dataset card declares the MIT license and states that copyright for the course materials belongs to Shkolkovo.online. It describes the intended use as LLM and STT/TTS training. That attribution is retained here. See UPSTREAM_README.md for the original card and NOTICE for provenance.

Please cite both the original dataset and this prepared release:

@misc{zeroagency_webinars,
  author = {ZeroAgency},
  title = {shkolkovo-bobr.video-webinars-audio},
  howpublished = {Hugging Face dataset},
  url = {https://huggingface.co/datasets/ZeroAgency/shkolkovo-bobr.video-webinars-audio}
}

@misc{teratts_webinars_2026,
  author = {TeraTTS},
  title = {Webinars: segmented audio with ASR transcripts},
  year = {2026},
  howpublished = {Hugging Face dataset},
  url = {https://huggingface.co/datasets/TeraTTS/webinars}
}