artmelancholy/golos_mfa_punctuation_long

Golos MFA Punctuation (Long)

Long-form Russian speech derived from govnejri/golos_mfa_punctuation.

Purpose

Most public Russian STT corpora ship as short clips (a few seconds each). For benchmarking long-form transcription, VAD, punctuation, and streaming behavior, you want minutes-long audio with reliable word-level alignments.

This dataset builds those long clips by splicing groups of consecutive short clips together, inserting randomized silences between them, and shifting the original word timestamps onto the new combined timeline.

Source

Built from the source dataset's sharded data/ directory (e.g. data/train-00000-of-00034.parquet).

Schema

field	type	description
idx	int32	1-based row index
audio	Audio (16 kHz mono WAV)	spliced audio with randomized inter-clip silences. Optionally speechnorm- and/or loudnorm-filtered when those flags are passed at build time.
duration	float32	combined clip duration in seconds
words	list<{text: string, start: f32, end: f32}>	word timings shifted onto the combined timeline

How it was built

The build pipeline — which splices consecutive short clips, inserts randomized silences, and shifts word timestamps onto the combined timeline — lives in a separate repository:

https://github.com/artmelancholy/golos_mfa_punctuation_long

See that repo for the source code, CLI flags, and reproduction instructions.

Usage

Load with datasets

from datasets import load_dataset

ds = load_dataset("artmelancholy/golos_mfa_punctuation_long", split="train")
row = ds[0]
audio = row["audio"]  # {"array": np.ndarray, "sampling_rate": 16000, "path": ...}
print(audio["array"].shape, audio["sampling_rate"], row["duration"], row["words"][:3])

Licensing

This repository contains derivative audio of the source dataset govnejri/golos_mfa_punctuation; consult the source dataset card for its licensing terms before redistribution.