| ## Variable descriptions | |
| | Variable | Description | | |
| | :------------------ | :----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | | |
| | speech_id | `id` of speech in SWERIK.<br> | | |
| | protocol_id | `id` of protocol in SWERIK. | | |
| | speech_number | Speech number within a protocol. | | |
| | dates | Date(s) from SWERIK for the given protocol the speech was in. | | |
| | name | Speaker name | | |
| | person_id | `id` of the [speaker/person in SWERIK](https://github.com/swerik-project/riksdagen-persons/). | | |
| | speaker_id | Same as `person_id` | | |
| | riksdagen_id | `id` of the person in Riksdagen's own system (if available). | | |
| | party | Party membership(s) of the speaker over their career. | | |
| | district | The district the person represents (if an MOP) | | |
| | role | Role(s) the speaker has had over their career. | | |
| | gender | Gender. | | |
| | start_segment | Start time of the speech within the media file (filename can be found in `audio_file` variable). Obtained by running speaker diarization on top of fuzzy string matching in order to refine existing timestamps from fuzzy string matching (the values in `start_text_time`). This is generally the most accurate estimate of when the speech was given. Time in seconds. | | |
| | end_segment | End time of the speech within the media file (according to diarization method). | | |
| | duration_segment | Duration of speech (diarization method). | | |
| | text | Original protocol transcript of speech from SWERIK. | | |
| | text_normalized | Normalized version of original protocol in a format suitable for string matching with wav2vec2 speech to text output (i.e. lowercase, only alphanumerics, convert numbers to spelled out letters, no punctuation). | | |
| | transcription_w2v | The automatically transcribed text in the region between `start_segment` and `end_segment`. | | |
| | start_text_time | Estimated start time of speech with fuzzy string matching (generally less accurate than `start_segment`) | | |
| | end_text_time | Estimated end time of speech with fuzzy string matching. | | |
| | born | Birth date of speaker. | | |
| | dead | Death date of speaker. | | |
| | bleu_score | BLEU score between `text_normalized` and `transcription_w2v` as an indication of the overlap. | | |
| | overall_score | Complicated score for fuzzy string matching that needs to be explained in a paper. | | |
| | nr_speech_segments | How many different speakers were identified by diarization within the duration of the speech (often but not always false negatives) | | |
| | start_segment_same | Is the `start_segment` of this speech a duplicate of some other speech's estimated `start_segment`? About ~150 speeches have the same identified start time. Generally either one of the two is correct, or there was a failure case of diarization that created too broad a speech duration that covered parts of two speeches.<br> | | |
| | audio_file | File name of the media file in Riksdagen's media archive. | |