Balalaika-longform

Russian speech for long-form text-to-speech: 189.13 hours of audio in 6,380 segments, each 30 seconds to 15 minutes long.

Balalaika-longform is a corpus for training and evaluating text-to-speech (TTS) systems on sustained speech. It retains long, continuous segments with their transcripts so that models can learn beyond the short utterances common in TTS training. The recordings include audiobook readings, lectures, talks, and other spoken material from YouTube.

All 313 source videos were marked Creative Commons Attribution (CC BY) in the collected source metadata. Every sample carries its original video URL, title, source time interval, and licensing evidence. Source credits are also available in metadata/sources.jsonl; see Licensing and attribution below.

The release contains the v3.1 enhanced audio, losslessly packaged as WebDataset TAR shards. Each sample has a FLAC recording, a punctuated Russian transcript, and JSON metadata. Train, validation, and test follow the channel-disjoint partition used in the long-form TTS experiments.

Dataset at a glance

Property	Value
Language	Russian
Intended task	Long-form text-to-speech
Full corpus	6,380 segments / 189.13 hours
Segment length	30–900 seconds
Sources	313 YouTube videos from 41 source channels
Audio	FLAC, 16-bit PCM; original sample rate and channel count preserved
Sample rates	44.1 kHz or 48 kHz
Primary transcript	Punctuated, cased GigaAM-v3-e2e-CTC output
Alternative transcript	ROVER consensus, where available
Distribution format	WebDataset, approximately 1 GB per shard

Audio is predominantly stereo. Downmix and resample for your model during data loading; the downloadable audio has not been re-encoded or resampled for this release. The source recordings were already compressed by YouTube, so FLAC describes the released encoding, not the original recording quality.

Splits

Split	Segments	Hours
train	6,043	170.68
validation	120	8.50
test	193	9.07
excluded — separate archival configuration	24	0.88
Full corpus	6,380	189.13

Partitioning happens by source channel, before creating short training windows. A channel belongs to only one of train, validation, or test. The original dev partition is named validation on the Hub; paper_split retains its original name in the metadata. Channel labels are grouping identifiers, not verified speaker identities.

The default configuration contains train, validation, and test: 6,356 segments / 188.25 hours. The excluded configuration preserves 24 additional segments from held-out channels that did not pass the validation/test ASR-consistency threshold. They are kept separately for completeness and must not be added to training. The full-corpus totals and duration figure include these archival samples.

WebDataset layout

data/
  train/train-00000.tar
  train/...
  validation/validation-00000.tar
  test/test-00000.tar
  excluded/excluded-00000.tar
metadata/
  sources.jsonl             # one provenance / attribution record per source video
  segments.jsonl            # lightweight sample index, splits, offsets, audio hashes
  shards.json               # exact shard paths, sample counts, sizes, checksums
  statistics.json
  release.json
assets/duration_distribution.png
SHA256SUMS
LICENSE.md

Each TAR shard contains successive groups of three files with a shared key:

<key>.flac     # enhanced speech segment
<key>.txt      # punctuated UTF-8 Russian transcript
<key>.json     # transcript alternatives, audio metadata, split, and source attribution

The key is the original sample identifier prefixed with balalaika_, with decimal points replaced by p for WebDataset compatibility. The unchanged original identifier is stored as id in JSON.

Loading the data

Hugging Face Datasets

pip install "datasets[audio]" soundfile

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

dataset = load_dataset(
    "lab260/Balalaika-longform", name="default", split="train", streaming=True
)
# Keep the FLAC bytes; decode them with soundfile below.
dataset = dataset.cast_column("flac", Audio(decode=False))
sample = next(iter(dataset))
waveform, sample_rate = sf.read(io.BytesIO(sample["flac"]["bytes"]), always_2d=True)
text = sample["txt"]
metadata = sample["json"]
mono = waveform.mean(axis=1)
print(sample_rate, mono.shape, text[:100], metadata["source"]["url"])

WebDataset

This example downloads one shard and reads its samples sequentially. Use the full list in metadata/shards.json for a complete training run.

pip install webdataset huggingface_hub soundfile

import io
import json
import soundfile as sf
import webdataset as wds
from huggingface_hub import hf_hub_download

shard = hf_hub_download(
    repo_id="lab260/Balalaika-longform",
    repo_type="dataset",
    filename="data/train/train-00000.tar",
)
dataset = wds.WebDataset(shard, shardshuffle=False)
for sample in dataset:
    audio, sample_rate = sf.read(io.BytesIO(sample["flac"]), always_2d=True)
    text = sample["txt"].decode("utf-8")
    metadata = json.loads(sample["json"])
    print(audio.shape, sample_rate, text[:100], metadata["source"]["url"])
    break

For TTS training, use .txt or the identical JSON text field. Adapt the sample rate to your model and use duration-aware batching to handle the long recordings. If you derive short windows, keep every window in its parent's split and align its transcript to the selected audio interval.

Metadata

Field	Meaning
id, key	Original sample identifier and WebDataset key
text	Primary punctuated transcript; identical to .txt
text_rover	Alternative multi-ASR consensus transcript, where available
transcript_source	Recognizer used for the primary transcript
duration_sec, sample_rate, channels, num_frames	Properties read from the released FLAC
segment_duration_sec	Rounded duration in the original corpus manifest
split, paper_split	Release partition and original experiment partition
speaker_id	Video-local diarization identifier, scoped by channel and video
source_start_sec, source_end_sec	Segment boundaries on the source video timeline
source	Video URL and title, channel information, and original CC-license evidence
audio_processing, audio_sha256	Processing description and checksum of the released audio

The original source metadata does not contain a resolved YouTube channel ID/name for every video. Missing values remain null; the catalog channel label and the direct video URL are retained. Speaker IDs are automatic diarization labels and should not be treated as globally verified identities.

Preparation and annotation

The corpus was assembled using the Balalaika speech-data pipeline. Source discovery selected videos marked Creative Commons. Diarization and context-aware segmentation produced long speech segments; duration selection and deduplication retained one segmentation per source video. DeepFilterNet3 speech enhancement produced the v3.1 audio distributed here.

Transcripts are automatic ASR annotations, not manually corrected ground truth. The primary text preserves punctuation and capitalization. text_rover preserves the available recognizer-voting transcript used in the experiments. These are different ASR outputs and can differ in words as well as punctuation.

The data was selected as single-speaker material by the collection pipeline, but automatic segmentation, speaker assignment, and transcripts can contain errors. Recordings may retain background sounds, reverberation, disfluencies, or additional voices. These are natural recordings rather than uniformly controlled studio speech. No corpus-wide DistillMOS threshold was applied to this release.

Licensing and attribution

All source videos in this release were marked Creative Commons Attribution on YouTube in the archived metadata. This is supported by a catalog cc: true flag, a saved YouTube Creative Commons Attribution license (reuse allowed) value, or both. These records were checked for all 313 included source videos when building the release; no conflicting or unknown license flag was accepted.

The source recordings retain their source licenses. Preserve the original video title, creator/channel attribution where available, source link, and licensing information when redistributing samples, and indicate the segmentation and speech enhancement applied here. The per-sample source object and source index provide this provenance. See YouTube's Creative Commons documentation and LICENSE.md.

The archived flags identify the CC BY license family without a version number. This release therefore preserves the source declaration instead of assigning a new CC version to every recording. The uploader-provided flag is licensing evidence, not an independent rights audit of every underlying text or performance.

For attribution corrections or removal requests, open a dataset discussion and include the relevant video URL or sample ID.

Acknowledgments

Credit belongs to the original creators and speakers linked in each sample. This corpus uses the Balalaika pipeline, GigaAM transcription models, recognizer-output voting (ROVER), and DeepFilterNet3 speech enhancement. Please cite the corpus URL and the exact dataset revision when reporting results.