RixVox-v2: A Swedish parliamentary speech dataset
RixVox-v2 is a parliamentary speech dataset spanning nearly 23000 hours of speech. The dataset was built by matching and force aligning speeches in parliamentary protocols to media recordings of debates. Each observation contains metadata about the speaker's name, gender, district, role, party affiliation, and the date the speech was given. We include identifiers for protocols, speeches and speakers that allow linking observations in our dataset to further metadata in SWERIK and the Riksdag's open data.
Variable | Description |
---|---|
text | Text from parliamentary protocols which have been force aligned with audio. Protocols from 1966-2002 sourced from SWERIK. Speeches from 2000-2024 sourced from the Riksdag's open data. |
audio | 16 kHz, mono audio segments, up to 30 seconds in duration. |
name | Speaker name. |
party | Party membership(s) of the speaker over their career. |
gender | Gender. |
role | Role(s) or position(s) the speaker has had over their career. |
district | The district(s) the person has represented over their career (available only if person has been a member of parliament). |
year | The year the speech was given according to parliamentary protocol metadata. |
dates | Date(s) when speech was possibly given according to parsed protocol metadata. A source protocol document can contain several debates spanning multiples dates. |
date_approx | Earliest possible date speech could have been given according to dates . Included to simplify date based filtering. |
start | Start time of this audio segment in the debate's audio file (seconds). See audio_file for source media filename. |
end | End time of this audio segment in debate audio file (seconds) |
duration | Duration of audio segment in seconds. |
audio_file | Filename of the source media file in the Riksdag's media archive. |
text_normalized | Normalized version of original protocol text used for calculating eval metrics (lowercased, alphanumeric, converted numbers to spelled out letters, no punctuation). |
text_timestamps | Text with timestamps inserted at sentence boundaries. Enables training Whisper with ability to output timestamps. |
text_previous | Text of the previous audio segment. Enables training Whisper with previous context as prompt. |
wav2vec_transcription | Audio transcribed by wav2vec2. |
whisper_transcription | Audio transcribed by Whisper-large-v3. |
bleu_whisper | BLEU score between text_normalized and normalized whisper_transcription . |
bleu_wav2vec | BLEU score between text_normalized and normalized wav2vec_transcription . |
wer_whisper | WER between text_normalized and normalized whisper_transcription . |
wer_wav2vec | WER between text_normalized and normalized wav2vec_transcription . |
cer_whisper_first | CER of 10 first characters of text_normalized and normalized whisper_transcription . |
cer_wav2vec_first | CER of 10 first characters of text_normalized and normalized wav2vec_transcription . |
cer_whisper_last | CER of 10 last characters of text_normalized and normalized whisper_transcription . |
cer_wav2vec_last | CER of 10 last characters of text_normalized and normalized wav2vec_transcription . |
is_silence | Whether the segment contains no text and no detected speech by wav2vec |
lang_prob_sv | Probability of spoken language being Swedish per the utput of whisper-large-v3's language detection. |
shard | Shard number. Corresponds to the shard numbering used in the naming of the uploaded parquet files. |
speaker_id | id of the speaker/person in SWERIK or in the Riksdag's systems. SWERIK id s start with i- . |
riksdagen_id | id of the person in the Riksdag's own system (if available). |
protocol_id | id of protocol/debate in SWERIK or the Riksdag's open data. id s from SWERIK start with prot- . |
speech_id | id of speech in SWERIK or the Riksdag's open data. id s from SWERIK start with i- |