---
title: 'SloPalSpeech: Slovak Parliamentary ASR Dataset'
url: https://www.emergentmind.com/topics/slopalspeech
type: topic
---

# SloPalSpeech: Slovak Parliamentary ASR Dataset

Searching arXiv for the specified corpus and closely related methodological work to ground the article with current paper metadata and citations.
SloPalSpeech is a publicly released large-scale Slovak automatic speech recognition dataset built from public parliamentary proceedings of the Slovak National Council. It was introduced to address the scarcity of Slovak ASR training data by converting long-form parliamentary recordings and official transcripts into aligned audio-text pairs suitable for model training. The released corpus contains 2,806 hours of speech segmented into 402,966 audio-text pairs, each no longer than 30 seconds, together with a full collection of segmented transcripts comprising about 60 million words and several fine-tuned Whisper models [2509.19270].

## 1. Corpus identity and motivation

SloPalSpeech is positioned as a response to a central problem in Slovak ASR: Slovak is treated as a low-resource language, and existing open datasets such as Common Voice, FLEURS, and VoxPopuli together provide only about 100 hours of Slovak speech. The corpus therefore aims to move Slovak beyond that low-resource regime by exploiting a source that is abundant, public, and regularly transcribed: parliamentary recordings and verbatim transcripts [2509.19270].

The source domain is parliamentary speech, specifically Slovak parliamentary hearings. The project advances Slovak ASR in three stated ways. First, it expands available Slovak speech data from roughly 100 open hours to 2.8k hours. Second, it introduces a practical pipeline for aligning long parliamentary recordings with noisy, imperfect transcripts and turning them into Whisper-compatible 30-second training examples. Third, it shows that the resulting data materially improves Slovak ASR across several Whisper model sizes.

The paper frames the corpus not merely as a dataset release but as evidence that high-volume public governmental archives can support modern ASR development in languages with limited open supervision. It explicitly states that the work “helps move Slovak ASR beyond low-resource status” [2509.19270].

## 2. Sources, composition, and released data

SloPalSpeech is derived from two official sources: MediaPortál (NRSR) for audio recordings of parliamentary hearings, and The Joint Czech and Slovak Digital Parliamentary Library for official full-text transcripts in DOCX format. The authors downloaded all available audio recordings by extracting HLS stream URLs and using FFmpeg, and they downloaded all available transcript DOCX files, discarding pages that redirected to HTML instead of valid DOCX [2509.19270].

The released resource is smaller than the raw archive because only the aligned and filtered portion enters the final ASR corpus. The paper also notes that the 60 million-word transcript release is a superset of the text actually covered by aligned audio, because alignment could not recover every transcript-audio match.

| Statistic | Value |
|---|---:|
| Raw collected audio | 4512.5 hours |
| Transcript files | 2111 |
| Successfully matched recordings | 1136 |
| Final aligned ASR dataset | 2806 hours |
| Segments after final WER filtering | 402,966 |
| Full segmented transcript release | ~60 million words |
| Clip duration target | up to 30 seconds |

The paper also reports intermediate filtering statistics. Before the final WER filter, the corpus contained 2977.73 hours and 427,276 segments. After filtering, it contained 2806.00 hours and 402,966 segments, meaning that 171.73 hours and 24,310 segments were removed.

Some corpus properties are left unspecified. The excerpt does not provide a train/dev/test hour breakdown, exact transcript-length statistics per segment, exact unique speaker counts, or exact split sizes in hours or examples. It does state that evaluation during fine-tuning used a held-out parliamentary test set of roughly 300 samples, a 300-sample subset of Slovak FLEURS, and a 300-sample subset of Slovak Common Voice [2509.19270].

## 3. Alignment, normalization, and segmentation pipeline

The technical core of SloPalSpeech is a long-form alignment pipeline designed for recordings ranging from seconds to 25 hours, with an average duration of 3.6 hours. Standard forced alignment was judged unsuitable because transcripts often contained text not actually spoken, recordings were too long for most forced-aligners’ sweet spot, and transcript/audio mismatch made strict forced alignment brittle [2509.19270].

Transcript processing begins with DOCX-to-XHTML conversion using Apache Tika Server, preserving formatting such as `<b></b>` tags. Bold lines were classified as speaker annotations if three conditions held: the text was bold, it contained at least one and no more than three names, and it did not exceed 15 words. To support this, the authors scraped a complete list of Slovak National Council members across electoral terms and compiled a set of known first and last names. They also removed all text inside `()` and `[]`, treating these as transcriber notes rather than spoken content, and added a heuristic for malformed notes in which a bracket started the note and a period ended it.

Before large-scale alignment, transcript validity was checked on five randomly sampled sessions. Initial WER values were greater than 1.0, which exposed a systematic mismatch: audio recordings contained only the forenoon part of sessions, while transcripts included both forenoon and afternoon proceedings. After removing afternoon portions from transcripts, the reported WERs became 12.9% for `openai/whisper-large-v3`, 21.7% for `openai/whisper-medium`, 36.1% for `openai/whisper-small`, and 77.3% for `openai/whisper-tiny`.

Alignment then proceeds in four stages. First, WhisperX is used to generate a reference transcript from the audio with word-level timestamps. Second, the reference transcript is aligned to the official transcript using anchors: words present in both transcripts. Candidate matches must be within Levenshtein distance 1, must not jump backward relative to the last aligned word, and must stay within a 50-word forward window. Additional heuristics skip words shorter than 3 characters and require a minimum match score of 3, where the score is based on agreement in the four preceding and four following words.

Third, 30-second candidate segments are constructed from anchors. The target is up to 30 seconds, but the implementation uses a soft threshold of 28 seconds plus a 2-second buffer. The procedure iterates through anchors, scans forward until finding the next anchor whose timestamp difference is greater than 28 seconds, then uses the ground-truth words between those anchors as the candidate transcript and extracts the corresponding audio slice.

Fourth, a final quality filter is applied by re-running Whisper directly on each segmented chunk shorter than 30 seconds and computing WER against the segment text. Based on empirical inspection, segments with WER < 40% were retained and segments with WER ≥ 40% were removed [2509.19270].

## 4. Fine-tuning experiments and benchmark results

SloPalSpeech was used to fine-tune four Whisper variants: Whisper-small, Whisper-medium, Whisper-large-v3-turbo, and Whisper-large-v3. Training used multiple NVIDIA A10 GPUs, Hugging Face Transformers, up to 3 epochs, and early stopping. For Whisper-small, Whisper-medium, and Whisper-large-v3-turbo, single-GPU fine-tuning was feasible. For Whisper-large-v3, the authors used Fully Sharded Data Parallel with 3 GPUs for training and 1 fourth GPU for inference and evaluation at selected checkpoints, ultimately stabilizing training with Transformers version 4.52.4 [2509.19270].

The learning-rate strategy differed by model scale. For small, medium, and turbo, the implementation followed a Hugging Face Whisper fine-tuning recipe inspired by Sanchit Gandhi’s blog, with the learning rate set 40× smaller than the pretraining learning rate rather than the common default of \(1 \times 10^{-5}\). For large-v3, that strategy led to rapid overfitting, so the final successful setup used learning rate \(1 \times 10^{-5}\) and weight decay \(0.01\).

The central result is that all four fine-tuned models improved on Slovak Common Voice 21 and FLEURS, with especially large relative gains for smaller models.

| Model | CV21 WER | FLEURS WER |
|---|---:|---:|
| Small-Base | 58.4 | 36.1 |
| Small-Tuned | 25.7 | 10.6 |
| Medium-Base | 38.0 | 18.7 |
| Medium-Tuned | 18.0 | 7.6 |
| Turbo-v3-Base | 31.7 | 10.7 |
| Turbo-v3-Tuned | 13.2 | 6.4 |
| Large-v3-Base | 20.8 | 9.2 |
| Large-v3-Tuned | 11.6 | 5.5 |

For Whisper-small, the relative improvement is reported as about 56% on Common Voice 21 and about 71% on FLEURS, which is the basis for the claim that its WER dropped by up to 70%. The paper also emphasizes that fine-tuned Whisper-small approaches baseline Whisper-large-v3 performance: on FLEURS, 10.6 versus 9.2; on Common Voice 21, 25.7 versus 20.8 [2509.19270].

The authors also highlight Whisper-large-v3-turbo as a practical efficiency-accuracy compromise, describing it as having roughly 730 million fewer parameters than large-v3, only about 100 million more than medium, and being within about 1% of large-v3 accuracy.

## 5. Relation to other parliamentary speech resources and naming ambiguities

SloPalSpeech belongs to a broader class of parliamentary speech-text corpora derived from public recordings and official proceedings. A closely related methodological reference is ParlaSpeech, which uses ParlaMint transcripts and public parliamentary recordings to build aligned corpora for Croatian, Polish, and Serbian, producing more than 5,000 hours of speech and accompanying text transcripts [2409.15397]. The relation is methodological rather than identificatory: ParlaSpeech is not SloPalSpeech, and it does not report a Slovak corpus under that name.

The distinction matters because the name can suggest multiple geographic or project associations. SloPalSpeech is a Slovak corpus built from the Slovak National Council. It is not a Slovenian or Slovene parliamentary corpus. The ParlaSpeech paper is especially relevant as a scalable parliamentary alignment framework, but its pilot languages are Croatian, Polish, and Serbian rather than Slovak [2409.15397].

A separate ambiguity arises from similarity to “SlothSpeech,” which is an ASR denial-of-service attack exploiting dynamic decoding latency rather than a parliamentary corpus or dataset release [2306.00794]. The two works are unrelated in purpose: SloPalSpeech addresses corpus construction and ASR adaptation, whereas SlothSpeech studies adversarial efficiency degradation.

This suggests that SloPalSpeech is best situated among parliamentary speech-resource construction efforts rather than among spoken dialogue or adversarial ASR work. Its distinctive contribution within that space is the Slovak focus and the demonstration that parliamentary archives can sustain successful Whisper fine-tuning.

## 6. Limitations, bias, and broader implications

The paper is explicit about several limitations. The most immediate is domain bias: because the data comes from parliament, the resulting models can become biased toward parliamentary language. The authors note that base Whisper often hallucinates generic repetitions like “Ďakujem,” whereas the fine-tuned models sometimes hallucinate parliamentary phrases such as “Pán poslanec” [2509.19270].

A second limitation is substitution of unknown words with parliament-related terms. The paper gives a concrete Common Voice example in which the reference “ostrov Nihau” is predicted as “ostrov Mihál,” where “Mihál” is a former politician’s surname. The authors also note general Whisper hallucinations and recommend an inference strategy like Faster-Whisper’s, where compression-ratio checks can trigger a re-decode with adjusted temperature.

A third limitation is multilingual degradation. Because all model parameters were fine-tuned, the models showed a strong shift toward Slovak and degraded English transcription performance. The excerpt does not quantify this loss, but it is explicitly observed. A fourth limitation is incomplete alignment coverage: the final 2.8k-hour dataset includes only the portion of audio that could be successfully aligned with transcripts. The full transcript collection is larger than the aligned ASR subset.

These limitations do not negate the corpus’s significance, but they define its scope. SloPalSpeech is a domain-specific Slovak ASR resource built from imperfect public archives, not a neutral general-speech corpus. Its broader implication is methodological: governments that publish full recordings and official transcripts may offer a practical route for building substantial ASR resources in other low-resource languages. A plausible implication is that parliamentary archives can function as a repeatable source of high-volume supervision when paired with sufficiently robust alignment and filtering procedures.

Source: https://www.emergentmind.com/topics/slopalspeech