Danish ASR Unified: 3.5 Million Speech Samples from 7 Sources

A unified Danish speech recognition dataset combines approximately 3.5 million audio samples from seven distinct sources, totaling roughly 16,000 hours of speech. The collection includes European and Danish Parliament recordings, read-aloud and conversational speech, broadcast media, and crowd-sourced samples. It was created by syvai and last updated on the Hugging Face platform in April 2026.

Use Cases

• Train Danish speech recognition models based on the large-scale, multi-source audio collection.
• Fine-tune models for parliamentary speech recognition based on the VoxPopuli and ftspeech sources.
• Develop models for conversational speech understanding based on the CoRal-v3 conversation subset.
• Benchmark ASR performance on read-aloud speech based on the CoRal-v3 read_aloud and nst-da sources.
• Create systems for transcribing broadcast media based on the nota source.

Strengths

• Large scale with approximately 3.5 million audio samples.
• Diverse speech sources covering parliamentary, conversational, read-aloud, and broadcast contexts.
• Significant total duration of roughly 16,000 hours of Danish speech.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Specific file formats, licensing, and detailed sample structure are unknown from the provided metadata.

Provenance

Source

Combined from 7 sources: VoxPopuli, ftspeech, CoRal-v3 read_aloud, nst-da, CoRal-v3 conversation, nota, Common Voice 17.

Collection Method

Likely aggregated and unified from existing public speech datasets.

Freshness

Last updated 2026-04-06 13:10:45; freshness should be verified.

Geography

Denmark (Danish language focus).