Datasets:

KlangAI
/

klang-dialects

license: cc-by-4.0
pretty_name: Klang Dialects
language:
  - sv
task_categories:
  - automatic-speech-recognition
size_categories:
  - 1K<n<10K
tags:
  - audio
  - speech
  - evaluation
  - dialects
configs:
  - config_name: sv-clean
    default: true
    data_files:
      - split: test
        path: data/sv/clean/test-*.parquet
  - config_name: sv-raw
    data_files:
      - split: test
        path:
          - data/sv/clean/test-*.parquet
          - data/sv/raw-only/test-*.parquet

Klang Dialects

Klang Dialects is an open benchmark for Swedish speech recognition, created by the Klang Research Team from recordings contributed through Knäck Klang.

The Swedish benchmark contains 1,804 recordings, 656 speaker IDs, and 5.15 hours of speech in its main configuration, sv-clean. The benchmark supports research on regional variation in Swedish speech recognition. We plan to expand the dataset with more sentences, speakers, and languages.

Configurations

Configuration	Recordings	Speaker IDs	Hours
`sv-clean` (default)	1,804	656	5.15
`sv-raw`	1,883	658	5.43

sv-clean provides high-quality reference transcripts for standard word error rate (WER) comparisons.

sv-raw includes all of sv-clean plus 79 recordings with more than one reasonable reference transcription. These recordings remain useful for benchmarking, although even a perfect recognizer may score above 0% WER on them.

Both configurations use a single test split and the same seven fields.

Loading

from datasets import load_dataset

clean = load_dataset(
    "KlangAI/klang-dialects", "sv-clean", split="test", revision="v1.0"
)
raw = load_dataset(
    "KlangAI/klang-dialects", "sv-raw", split="test", revision="v1.0"
)

Fields

Field	Type	Description
`id`	string	Unique recording ID.
`audio`	Audio	Mono speech recording in 16-bit FLAC at 16 kHz.
`text`	string	Reference transcript, including capitalization and punctuation.
`speaker`	string	Anonymous speaker ID. A person who contributed more than once may have multiple IDs.
`region`	string or null	Contributor-supplied three-digit Swedish postcode prefix, where available.
`duration`	float64	Recording length in seconds.
`prompt_level`	int64	Difficulty of the original prompt: 1 (easy), 2 (medium), 3 (hard).

Audio is included directly in the Parquet files.

Collection

Knäck Klang invited participants to challenge Klang's Swedish speech recognition: the more transcription errors, the higher their score. Participants joined mainly through social media and could compare the transcription with the sentences on screen.

The challenge used a bank of 111 prompts, ranging from everyday language to longer constructions and tongue-twisters. Each participant read one randomly selected prompt at each of three difficulty levels.

The collection took place during the first half of 2026. Contributors opted in to an open speech dataset through the collection website. Geographic coverage is strongest in southern Sweden, with contributions from across the country.

Reference quality

Participants sometimes skipped or replaced words as they read. The Klang Research Team prepared the references from the recordings so that these changes would not be counted as recognition errors.

The team compared transcripts from multiple ASR models with the prompts and used alignment checks to identify discrepancies. The team manually reviewed roughly 1,000 recordings, confirming or correcting references by listening. References were also accepted automatically when multiple ASR models agreed with the prompt and passed the quality checks.

The final dataset contains 259 distinct reference texts in sv-clean and 319 in sv-raw. In sv-clean, 203 recordings have references that differ from their assigned prompt. Examples include:

Sentence on screen	Reviewed transcript
Jag glömde köpa mjölk kan du svänga förbi butiken?	Jag glömde köpa mjölk kan du svänga förbi affären?
Vi såg en riktigt bra film på bio förra veckan.	Vi såg en riktigt bra film förra veckan.

In the second example, an accurate transcription would score 20% WER against the ten-word prompt because the speaker omitted two words.

License and contact

The dataset is available under Creative Commons Attribution 4.0 International.

For questions, corrections, or a recording removal request, contact Klang.

Citation

Collected and maintained by the Research team at Klang.

@misc{klang_dialects_2026,
  author = {{Klang}},
  title = {Klang Dialects},
  year = {2026},
  url = {https://huggingface.co/datasets/KlangAI/klang-dialects}
}