NbAiLab/nb-asr-qwen3whisperxagreement-v1
nb-asr-qwen3whisperxagreement-v1 Word-level forced alignment training data for Norwegian speech, produced by keeping only examples where two independent aligners — WhisperX and Qwen3 (Lunde forced aligner) — agree within a tight tolerance. Dataset Description This dataset contains 702,067 speech segments drawn from the NB-ASR Norwegian audio corpus. Each record pairs an audio file with a word-level forced alignment in a format suitable for training a… See the full description on the dataset page: https://huggingface.co/datasets/NbAiLab/nb-asr-qwen3whisperxagreement-v1.
nb-asr-qwen3whisperxagreement-v1
Word-level forced alignment training data for Norwegian speech, produced by keeping only examples where two independent aligners — WhisperX and Qwen3 (Lunde forced aligner) — agree within a tight tolerance.
Dataset Description
This dataset contains 702,067 speech segments drawn from the NB-ASR Norwegian audio corpus. Each record pairs an audio file with a word-level forced alignment in a format suitable for training a <timestamp>-predicting sequence model (e.g. Qwen3 fine-tuning).
Timestamps are represented as integer frame indices at 80 ms resolution (12.5 frames/sec), matching the feature-frame rate used during alignment.
How the data was created
Two forced-alignment systems were run independently on the same audio:
For each sample, word boundaries from the two systems were compared after converting both to 80 ms frame indices. A pair of alignments is kept only when every word satisfies:
|start_whisperx − start_qwen3| ≤ 2 frames(≤ 160 ms)|end_whisperx − end_qwen3| ≤ 5 frames(≤ 400 ms)
Samples with a different word count between the two systems are rejected outright.
Filtering outcome (from `filter_by_whisperx.py`):
~42.9 % of candidates passed the agreement filter.
Data Format
Each record is a JSON object with four fields:
Example record
{
"id": "DVNA20000613_1609520_1623520",
"audio": "/nird/datalake/NS30001K/nb-asr-audio/wav16mono/ad/31/DVNA20000613_1609520_1623520.wav",
"text": "Alle <timestamp> <timestamp> de <timestamp> <timestamp> fem <timestamp> <timestamp> ...",
"labels": [-100, 3, 6, -100, 6, 8, -100, 8, 11, ...]
}The text and labels fields align token-by-token:
token: Alle <timestamp> <timestamp> de <timestamp> <timestamp> ...
label: -100 3 6 -100 6 8 ...Frame index 3 corresponds to 3 × 0.08 s = 0.24 s.
Intended Use
This dataset is intended for supervised fine-tuning of models that predict word-level timestamps from audio, particularly those using a <timestamp> token scheme. The double-agreement filter provides higher-confidence ground-truth boundaries than using either aligner alone.
Source Audio
Audio originates from the NB-ASR Norwegian broadcast and parliamentary speech corpus. Files are 16 kHz mono WAV, stored at paths under /nird/datalake/NS30001K/nb-asr-audio/.