ParCzech4Speech 1.0: Czech Parliamentary Speech Corpus

• The paper introduces an innovative speech alignment methodology using WhisperX and Wav2Vec 2.0 to achieve precise, word-level timestamps.
• ParCzech4Speech 1.0 is a comprehensive resource offering three variants—sentence-segmented, unsegmented, and raw-alignment—tailored for diverse speech tasks.
• It integrates extensive metadata from ParCzech 4.0, enabling robust research in ASR, TTS, speaker modeling, and political discourse analysis.

ParCzech4Speech 1.0 is a processed Czech parliamentary speech corpus derived from ParCzech 4.0 and corresponding AudioPSP recordings of the Chamber of Deputies. It is designed for speech-oriented tasks, including automatic speech recognition, speech synthesis, alignment, and speaker modeling, and is released in three variants: sentence-segmented, unsegmented, and raw-alignment. Its largest variant contains 2,695 hours of aligned speech, and the corpus preserves the metadata structure of ParCzech 4.0 while adding automated audio-text alignment produced with WhisperX and Wav2Vec 2.0 (Stankov et al., 8 Sep 2025).

1. Corpus provenance and scope

ParCzech4Speech 1.0 is built from two primary sources: the ParCzech 4.0 parliamentary text corpus and the AudioPSP 24.01 collection of official sound recordings of Czech parliamentary sessions. ParCzech 4.0 provides the official stenographic transcripts, originally encoded in TEI XML, while AudioPSP supplies the corresponding multi-speaker session audio. The corpus targets modern speech processing tasks rather than purely textual corpus linguistics, which distinguishes it from ParCzech 4.0 despite their shared metadata backbone (Stankov et al., 8 Sep 2025).

The resource is also positioned as a methodological successor to the ParCzech 3.0 speech recognition version. The later pipeline replaces earlier GMM-HMM-style alignment with a neural workflow based on WhisperX and Wav2Vec 2.0, with the stated effect of extracting more usable speech and achieving higher alignment reliability. In this respect, ParCzech4Speech 1.0 belongs to a broader shift in Czech ASR from hybrid LVCSR pipelines toward self-supervised transformer-based systems, a transition documented for Czech by monolingual Wav2Vec 2.0 work on CommonVoice, VoxPopuli, and MALACH (Lehečka et al., 2022).

The corpus occupies an important place among Czech speech resources because the underlying data are large-scale parliamentary proceedings rather than read speech or narrow-domain recordings. The overview in the source description contrasts it with smaller, more restricted, or more specialized Czech corpora such as Vystadial, PDTSC 2.0, and domain-limited resources, while emphasizing permissive licensing and broad metadata retention (Stankov et al., 8 Sep 2025).

2. Source transformation and alignment workflow

The construction process begins by converting the ParCzech 4.0 TEI transcripts into TSV files aligned to AudioPSP recordings. The TSV representation is token-centric: one row per token, with fields such as token_id, token_text, speaker_id, date_of_speech, and token_type. The token_id remains compatible with ParCzech 4.0, enabling later reattachment of speaker role, gender, birth date, and related metadata (Stankov et al., 8 Sep 2025).

A central difficulty is that official parliamentary transcripts are not literal renderings of speech. They are grammatically normalized and often omit disfluencies, repetitions, and other spoken artifacts. Because of this, the transcripts cannot be used directly as a conventional forced-alignment target. The pipeline therefore proceeds in two stages. First, WhisperX performs VAD-based segmentation, Whisper ASR transcription, and Wav2Vec 2.0 CTC segmentation to obtain recognized words with word-level timestamps. Second, the recognized word sequence is aligned to the official ParCzech transcript via dynamic-programming-style word alignment (Stankov et al., 8 Sep 2025).

The alignment is formulated over official words $o_1, \dots, o_M$ and recognized words $r_1, \dots, r_N$. For each candidate pair $(o_j, r_i)$, the pipeline computes Levenshtein distance $d(o_j, r_i)$. If

$d(o_j, r_i) > |o_j|,$

the system does not force a match and instead inserts a gap token on the misaligned side. Otherwise, the recognized word is aligned to the official token and its timestamps are propagated to the official word. The result is a word-level table containing official words, recognized words, edit distances, timestamps where available, and identifiers linking the row both to ParCzech 4.0 and to AudioPSP recordings (Stankov et al., 8 Sep 2025).

The pipeline must also account for known timestamping limitations in the Wav2Vec 2.0 alignment stage. Digits such as 123 or 1,000 and pronounced special characters such as %, {, }, @, § may appear in the recognized transcript without timestamps. This limitation directly affects boundary inference, especially for segments whose edges depend on numeric tokens. To mitigate this, the corpus construction includes a re-recognition step with a Wav2Vec 2.0 model, followed by comparison against Whisper transcripts; the comparison maps numeric forms such as 1,000 to spoken forms such as tisíc to reduce superficial mismatches (Stankov et al., 8 Sep 2025).

3. Quality control, filtering logic, and reliability signals

ParCzech4Speech 1.0 does not treat all matched text-audio spans as equally reliable. Instead, it applies explicit filtering criteria at segment level. In the sentence-segmented variant, a segment must have timestamped boundary words, must be single-speaker, and is discarded if it contains gap words other than numbers or special symbols. Segments beginning or ending with untimestamped numeric or special-symbol tokens are also removed (Stankov et al., 8 Sep 2025).

Four automatic metrics are stored for sentence-segmented segments. The first is segment edit distance, defined as the Levenshtein distance between normalized official and recognized segment texts, with threshold

$\text{SegmentEditDistance} \leq 9.$

The second is maximum aligned edit distance, with threshold

$\max_j d(o_j, r_{\pi(j)}) \leq 5,$

which prevents any single aligned word from being too divergent. The third is average character duration,

$\bar{d}_c = \frac{T}{C},$

where $T$ is segment duration and $C$ is the number of characters, filtered to

$r_1, \dots, r_N$0

The fourth is a speaker-text counter used for analysis rather than filtering (Stankov et al., 8 Sep 2025).

The unsegmented variant applies related but not identical logic. Consecutive well-aligned official words are merged until a gap or speaker change is encountered. Quality thresholds again include segment edit distance $r_1, \dots, r_N$1 and maximum aligned edit distance $r_1, \dots, r_N$2, but character-duration filtering is applied per word: $r_1, \dots, r_N$3 with

$r_1, \dots, r_N$4

In addition, each segment must contain at least 5 original words, and segments longer than 30 seconds are randomly split while preserving the minimum-word constraint (Stankov et al., 8 Sep 2025).

These procedures yield several nested coverage levels. Table 1 of the source description reports 6,431 hours of source audio, 4,770 hours of recognized material, 3,308 hours of aligned material, and 2,990 hours of perfectly aligned material, with 27,835,729 aligned words and 25,562,587 perfectly aligned words. Speaker coverage remains high through filtering: 590 speakers in the source, 589 recognized, and 588 aligned or perfectly aligned (Stankov et al., 8 Sep 2025).

A broader methodological implication is that ParCzech4Speech 1.0 operationalizes alignment reliability through multiple weak signals rather than through a single ASR confidence score. This design is consistent with recent Czech speech processing work in which Wav2Vec 2.0 is used both as a recognition model and as an alignment-enabling backbone for parliamentary-like speech (Lehečka et al., 2022).

4. Released variants and quantitative profile

The corpus is distributed in three variants with distinct intended uses. The sentence-segmented variant is optimized for ASR and TTS with clean sentence-level boundaries defined by UDPipe sentence segmentation and word-level timestamps. The unsegmented variant preserves utterance flow across sentence boundaries and is intended for tasks such as streaming ASR or long-context modeling. The raw-alignment variant exposes the full word-level alignment table for downstream custom segmentation and filtering (Stankov et al., 8 Sep 2025).

Variant	Total duration	Structure
Sentence-segmented	$r_1, \dots, r_N$5 h	Sentence-bounded, single-speaker, filtered
Unsegmented	$r_1, \dots, r_N$6 h	Continuous utterance blocks, filtered
Raw-alignment	Not summarized by segment totals	Full word-level alignment tables

The sentence-segmented variant contains 682,254 training segments over 1,131 hours, 5,094 development segments over 10.14 hours, and 11,379 test segments over 20.63 hours. The corresponding word counts are 8,684,912 for train, 74,556 for dev, and 164,803 for test. Speaker coverage is 525 in train, 29 in dev, and 30 in test. The split design follows speaker-based partitioning inherited from ParCzech 3.0, although the test split lacks one speaker from the original ParCzech 3.0 test set (Stankov et al., 8 Sep 2025).

The unsegmented variant is substantially larger. It contains 1,311,027 training segments over 2,631 hours, 20,352 development segments over 43.43 hours, and 9,127 test segments over 21.37 hours. Word counts are 18,385,517 for train, 326,359 for dev, and 147,514 for test, for a total duration of about 2,695.8 hours. Speaker coverage is 527 in train and 30 in each of dev and test (Stankov et al., 8 Sep 2025).

The raw-alignment variant is structurally different from the other two releases. It is not a segmented corpus in the usual ASR sense, but a set of TSV files exposing official words, recognized words, word-level edit distance, timestamps, and identifiers. This variant is intended for users who wish to impose alternative alignment thresholds, recover longer spans, or design different segmentation criteria (Stankov et al., 8 Sep 2025).

5. Metadata architecture and relation to adjacent resources

A defining property of ParCzech4Speech 1.0 is that the speech-aligned data remain tightly connected to the richer ParCzech 4.0 textual corpus. Through token and segment identifiers, users can recover speaker identity, gender, date of birth, party affiliation, role, session date, session ID, topic information, and positions within the session. The speech pipeline further adds start and end times for matched official words, recognized-word strings, word-level edit distances, split assignments, and segment-level quality metrics such as segment_edit_distance, max_aligned_edit_distance, and character-duration measures (Stankov et al., 8 Sep 2025).

This architecture gives the resource a dual character. On one side, it is a training corpus for supervised speech tasks. On the other, it is an alignment layer over an already structured parliamentary corpus. That design brings it close to the broader family of spoken parliamentary corpora represented by ParlaSpeech. In ParlaSpeech 3.0, the Czech component, ParlaSpeech-CZ 1.0, provides 1,218 hours, 717,682 sentences, Universal Dependencies-style linguistic annotation, sentence-level sentiment labels, and automatic filled-pause detection, but it does not yet provide MFA-based word or grapheme alignment for Czech (Ljubešić et al., 3 Nov 2025). ParCzech4Speech 1.0 differs by prioritizing flexible speech-oriented releases, especially word-level alignment tables and two distinct segmentation regimes (Stankov et al., 8 Sep 2025).

The corpus also sits alongside Czech-specific alignment tooling such as Prak, which is an open-source Czech forced aligner generating phone sequences from text and aligning them to audio. Prak operates with a Czech pronunciation generator, a hybrid HMM-NN decoder, and Praat/TextGrid outputs, and is explicitly designed for phone-level time boundaries rather than parliamentary-scale text-audio matching (Hanžl et al., 2023). This suggests a complementary relationship: ParCzech4Speech 1.0 provides large-scale parliamentary alignment at the word and segment level, whereas Prak addresses a finer phonetic granularity.

On the ASR side, ParCzech4Speech 1.0 is technically adjacent to Czech Wav2Vec 2.0 work showing that monolingual self-supervised transformers can exploit large Czech unlabeled corpora, perform strongly on VoxPopuli parliamentary speech, and benefit from multi-domain followed by in-domain fine-tuning (Lehečka et al., 2022). Since VoxPopuli is itself parliamentary speech, the methodological overlap is direct at the domain level even though the corpus releases serve different purposes.

6. Applications, limitations, and future development

The sentence-segmented variant is suited to standard supervised ASR and TTS because it provides bounded, single-speaker units with explicit quality controls. The unsegmented variant better supports streaming ASR, long-context acoustic modeling, and discourse-level analysis because it preserves utterance flow across sentence boundaries. The raw-alignment release supports custom segmentation, alignment research, and alternative quality-selection strategies (Stankov et al., 8 Sep 2025).

The corpus is also designed for research beyond recognition. Its metadata support speaker modeling, role-conditioned analysis, political discourse studies, and temporal analysis across sessions. Because the data are parliamentary, topic, speaker, and session metadata can be combined with acoustic spans in a way that is difficult in generic speech datasets. The source description explicitly notes use cases in ASR training and evaluation, TTS, speaker modeling and diarization research, sociolinguistic and political speech analysis, and alignment and segmentation research (Stankov et al., 8 Sep 2025).

Several limitations follow directly from the construction method. The domain is parliamentary and therefore formal, institutional, and only partly spontaneous. Official transcripts are normalized and may omit disfluencies, partial words, and some spoken material, which constrains disfluency modeling and creates a systematic mismatch between text and acoustics. Numbers and special characters are weaker alignment points because Wav2Vec 2.0 timestamping does not cover them reliably. Audio quality reflects plenary-session conditions, including floor noise, applause, and occasional overlapping speech, although the sentence-segmented variant excludes multi-speaker segments (Stankov et al., 8 Sep 2025).

For Czech speech technology, these limitations connect ParCzech4Speech 1.0 to a broader linguistic issue: spoken Czech and formal orthography do not always coincide. Transformer-based ASR work on MALACH showed that Wav2Vec 2.0 can be trained to output formal Czech even when the underlying speech is colloquial, reducing dependence on explicit pronunciation lexicons and normalization rules (Lehečka et al., 2022). That result does not alter the corpus itself, but it clarifies why a resource based on normalized official parliamentary transcripts remains useful for end-to-end ASR.

The release is under CC-BY, allowing academic and commercial use subject to attribution. All three variants are distributed through the LINDAT/CLARIAH-CZ repository at http://hdl.handle.net/11234/1-5946, and the sentence-segmented and unsegmented variants are additionally available on Hugging Face as ufal/parczech4speech-segmented and ufal/parczech4speech-unsegmented (Stankov et al., 8 Sep 2025). The stated future directions are to fine-tune ASR models on ParCzech4Speech, improve alignment for numbers and special characters, extend coverage to additional parliamentary terms or sessions, and add further annotation layers such as prosody, disfluencies after manual correction, or additional syntactic and semantic tags (Stankov et al., 8 Sep 2025).