akisviete/azuolas-nonspeech-pool

Non-speech pool — Ąžuolas v2 training auxiliary

The non-speech clip pool used to train akisviete/azuolas-qwen-lt (Ąžuolas v2, Qwen3-ASR-1.7B Lithuanian). Injected during full SFT at nonspeech_rate 0.0142 (realised 1.40 % of the training stream); this is what buys the model's measured 100 % non-speech rejection (0 characters on silence / hiss / room tone / ESC-50 / DEMAND).

Stats: 19,755 clips · 42.57 h · 16 kHz mono WAV · all text empty.

Every clip was filtered with Silero VAD — any clip containing detected speech activity was removed, so the pool is verified non-speech (see also snakers4/silero-vad).

source	clips	hours	license
wavs/audioset/ (AudioSet non-speech)	10,564	~27.9 h	CC BY 4.0
wavs/synth_v1/ (generated)	4,699	~5.5 h	generated (no rights holder)
wavs/synth_v2/ (generated)	3,196	~5.4 h	generated (no rights holder)
wavs/musan/ (MUSAN noise)	1,296	~3.7 h	CC BY 4.0

Layout

data.jsonl          # one row per clip: audio (repo-relative), duration_s, source, license, text, target
wavs/<source>/*.wav # 16 kHz mono (audioset split into a/ + b/)

target is the training target for every clip: language None<asr_text> — i.e. the model is taught to answer "None" (no speech) and emit an empty transcription.

Usage

from datasets import load_dataset
ds = load_dataset("akisviete/azuolas-nonspeech-pool", data_files="data.jsonl",
                  base_path=".")  # after `huggingface-cli download --repo-type dataset`

Licence

Every clip is CC BY 4.0 or generated in-house (no rights holder). Attribution as per the source licences: MUSAN (CC BY 4.0, Johns Hopkins); AudioSet non-speech clips as curated (CC BY 4.0).

Questions: akisviete@gmail.com.