samuli/eduskunta-asr

Finnish Parliament ASR Dataset

Speech dataset from Finnish parliament (Eduskunta) plenary sessions with aligned ASR transcriptions and human reference texts.

Overview

	Train	Test
Segments	64,945	9,121
Duration	423 h	59 h
Speakers	212	187

Total: 74,066 segments, 482 hours, 212 unique speakers.

Audio is 16 kHz mono Opus. Segments are individual speaker turns. Segments with CER > 0.30 are excluded (see below).

Schema

Column	Type	Description
audio	Audio	Opus audio clip (16 kHz mono)
transcription	string	Human reference text from official parliamentary transcript (poytekirja)
asr	string	Whisper ASR output
cer	float	Character error rate between ASR and human text
speaker	string	Speaker name
party	string	Parliamentary group (sd, kok, kesk, ps, vihr, vas, kd, r, sin)
session	string	Session identifier (e.g. session_152_2018)
date	string	Session date
start_time	float	Segment start time in source audio (seconds)
end_time	float	Segment end time in source audio (seconds)
duration	float	Segment duration (seconds)
uuid	string	Unique segment identifier

Speakers and parties

213 unique speakers from the 2016-2018 parliamentary term. Party distribution reflects speaking time, not seat count.

Party	Code	Segments
Social Democrats	sd	17,368
Centre Party	kesk	13,640
National Coalition	kok	9,185
Finns Party	ps	8,230
Left Alliance	vas	6,831
Greens	vihr	5,646
Christian Democrats	kd	4,393
Swedish People's Party	r	2,640
Blue Reform	sin	2,383

~10,200 segments have no party label (ministers speaking in official capacity, speakers of parliament).

Character error rate (CER)

Each segment includes a cer field: the character-level edit distance between the Whisper ASR output (asr) and the official parliamentary transcript (transcription). Lower is better.

Note that the human reference texts are edited transcripts, not verbatim speech — filler words are removed, grammar is corrected, etc. The CER therefore reflects both ASR errors and editorial normalization.

Segments with CER > 0.30 are excluded from the dataset as they represent failed alignments (the ASR output mapped to the wrong section of the transcript). The remaining distribution:

CER range	Segments	Cumulative
0.00 - 0.05	12,616 (17%)	17%
0.05 - 0.10	24,895 (34%)	51%
0.10 - 0.15	18,420 (25%)	76%
0.15 - 0.20	10,862 (15%)	90%
0.20 - 0.30	7,273 (10%)	100%

For ASR fine-tuning, consider filtering further (e.g. CER <= 0.15 retains 76% of the data with higher alignment confidence).

Train/test split

Split by parliamentary session (90/10). All segments from a given session appear in the same split. 225 sessions in train, 25 in test.

Source data

• Audio: Plenary session recordings from verkkolahetys.eduskunta.fi
• Transcripts: Official parliamentary transcripts (poytakirjat) from the Eduskunta open data API
• Time period: May 2016 - February 2019 (250 sessions)

Note: The verkkolahetys.eduskunta.fi API used to collect this data has since been restructured. The pipeline description below documents the original approach in enough detail to reimplement against the current API.

Pipeline

1. Scrape session metadata

Fetch the list of plenary sessions from the Eduskunta video archive API (verkkolahetys.eduskunta.fi/api/v1/categories/slug/taysistunnot). For each session, fetch the official transcript (poytakirja) from verkkolahetys.eduskunta.fi/api/v1/eventmetas/transcripts/PTK+{session_number}%2F{year}+vp. The transcript JSON contains structured speech data: each agenda item (asiakohta) has speaker turns (puheenvuoro) with speaker name, party, content paragraphs, and timestamps. Also download the full session audio (MP4/WAV) from the media URL in the session metadata.

2. ASR with speaker diarization

Run WhisperX on the full session audio with --diarize and --language fi. This produces timestamped ASR segments, each labeled with a speaker ID from diarization (e.g. SPEAKER_07). The output is one JSON file per session with segments containing start, end, text, and speaker fields.

3. Forced alignment

Align ASR segments to the official transcript text using character-level edit distance (Levenshtein). This is the core of the pipeline — it maps each ASR segment to the corresponding span in the human transcript, enabling extraction of (audio, reference text) pairs.

Text normalization: Before comparison, both ASR and transcript text are normalized by removing punctuation, converting to lowercase, and collapsing whitespace.

Two-phase search for each ASR segment:

1. Coarse region finding: The transcript is treated as a word array. A sliding window (sized to match the ASR segment word count) steps through the transcript words, computing CER at each position. Search starts from the end position of the previous segment and expands outward with forward priority (5 forward steps per 1 backward step). If a region with CER <= 0.30 is found, it's returned immediately. Otherwise, the top 3 candidate positions are returned.

2. Fine boundary tuning: For each candidate region, try all combinations of start offset (+/- 15 words) and window length (+/- 15 words from the ASR segment length). Compute CER for each candidate span. Keep the span with the lowest CER. Early-stop once CER drops below 0.05 and starts rising again.

Speaker-aware alignment: The transcript is split by speaker turns. Each ASR segment is first matched within nearby speaker segments (up to 3 candidates based on word position proximity). If no good match is found within speaker boundaries, a global fallback searches the full transcript. Speaker metadata (name, party, agenda item number) is attached from whichever transcript segment the match lands in.

CER computation: Levenshtein.distance(asr_normalized, human_normalized) / max(len(asr_normalized), len(human_normalized)). Using max length avoids bias toward shorter texts.

4. Audio clipping

For each aligned segment, extract the audio between start_time and end_time (with 10ms padding and fade in/out) from the full session audio. Export as Opus at 16 kHz mono. Each clip is named by UUID.

5. Quality filtering

Segments with CER > 0.30 are excluded. These are typically failed alignments where the ASR output matched the wrong part of the transcript (e.g. due to long silence gaps, overlapping speech, or procedural segments not present in the transcript).

License

CC-BY-4.0.

Audio source: verkkolahetys.eduskunta.fi. At the time of collection, the service's terms permitted free distribution and use with attribution. Please verify the current terms before redistribution.

Citation

If you use this dataset, please cite:

@misc{eduskunta-asr-2025,
  title={Finnish Parliament ASR Dataset},
  author={Samuli Sillanpää},
  year={2025},
  url={https://huggingface.co/datasets/samuli/eduskunta-asr}
}