README.md

---
language:
- sv
license: cc-by-4.0
pretty_name: Swedia ASR Dataset
task_categories:
- automatic-speech-recognition
tags:
- swedish
- asr
- speech-recognition
- dialects
- transcription
- wer
- cer
---

# Swedia ASR Dataset

This repository contains a small Swedish ASR evaluation dataset based on speech
transcriptions from Swedia 2000. It was assembled to compare automatic
speech-recognition output against manually corrected reference transcriptions
for Swedish dialectal speech.

The dataset is useful for quick experiments with Swedish ASR systems, especially
when you want to inspect recognition quality on spontaneous speech from
different regions, speakers, ages, and genders.

## Thesis Context

This dataset was prepared as part of Kajsa Vesterberg's bachelor thesis,
[*Lost in Translation: AI's Struggles with Scanian - A Study on Language Models'
Attempts to Conquer Swedish and its Dialects*](https://lup.lub.lu.se/student-papers/search/publication/9191345),
published through Lund University's LUP Student Papers repository in 2025.

The thesis investigates whether automatic speech recognition performs unevenly
across Swedish regional speech, with particular attention to Scanian dialects.

## Repository Contents

| File | Description |
| --- | --- |
| `swedia_asr_dataset.tsv` | Normalized row-level TSV with reference transcripts, ASR outputs, and computed WER/CER. |
| `Dataset-Swedia-2000.docx` | ASR/dictation output grouped by location and speaker category. |
| `Original-Texts-Swedia-2000.docx` | Manually corrected reference transcriptions for the same speech excerpts. |
| `Calculation-of-Results-Swedia-2000.xlsx` | Spreadsheet with WER/CER calculations and comparison notes. |

## TSV Format

`swedia_asr_dataset.tsv` is the easiest file to use programmatically. It has 56
rows, one row per location and speaker-group excerpt.

Columns:

- `location`
- `speaker_group`
- `reference_transcript`
- `word_diktering_transcript`
- `word_diktering_wer`
- `word_diktering_cer`
- `whisper_transcript`
- `whisper_wer`
- `whisper_cer`

Some `word_diktering_*` fields are empty because the Word Dictation section does
not contain every location/speaker combination that appears in the reference and
Whisper sections.

The original DOCX and XLSX files are intentionally kept in the repository. They
represent the source working documents from the thesis project, while
`swedia_asr_dataset.tsv` is a cleaned, analysis-friendly version derived from
them.

## Speaker and Location Structure

The documents are organized by location and speaker group. Examples of locations
included in the files are:

- Bara
- Bjuv
- Broby
- Löderup
- N Rörum
- Össjö
- Kårsta
- Skuttunge
- Kyrkslätt, Finland
- Särna, Dalarna
- Sproge, Gotland

Speaker groups include:

- `Äldre Kvinna`
- `Äldre Man`
- `Yngre Kvinna`
- `Yngre Man`

## Metrics

The normalized TSV includes computed evaluation fields for:

- `WER`: word error rate
- `CER`: character error rate

The original spreadsheet also includes comparison fields for:

- comparisons across locations/regions
- comparisons across speaker age and gender groups
- Word Dictation vs. Whisper-style ASR comparisons

## Suggested Uses

- Evaluate Swedish ASR output against reference transcripts.
- Compare recognition quality across dialect regions.
- Inspect typical ASR errors in spontaneous Swedish speech.
- Practice WER/CER calculation and ASR error analysis.
- Load a single TSV instead of manually extracting the DOCX files.

## Loading the Files

For most analysis, start with the normalized TSV:

Example:

```python
import pandas as pd

data = pd.read_csv(
    "hf://datasets/kvest/Swedia-ASR-Dataset/swedia_asr_dataset.tsv",
    sep="\t",
)
print(data[["location", "speaker_group", "whisper_wer", "whisper_cer"]].head())
```

The original DOCX/XLSX files are still included for transparency and manual
inspection.

## Limitations

- The dataset is small and best suited for exploratory ASR evaluation.
- The repository contains transcriptions and evaluation material, not raw audio.
- Results should be interpreted as a compact benchmark/sample rather than a
  comprehensive Swedish ASR evaluation suite.

## Citation and Source Notes

This dataset is based on transcribed material from Swedia 2000 and was prepared
for ASR comparison and error analysis in the bachelor thesis linked above. If
you use it, cite this dataset page, the thesis, and Swedia 2000 as the
underlying transcription source where appropriate.