EuroSpeech: A Multilingual Speech Corpus

TL;DR

EuroSpeech introduces a scalable, open-source pipeline for constructing multilingual speech datasets from parliamentary proceedings and presents a large-scale EuroSpeech corpus across 22 European languages. The pipeline features robust data sourcing, a two-stage dynamic alignment to handle non-verbatim transcripts, and CER-based filtering to produce high-quality audio-text pairs. Finetuning a pretrained multilingual ASR model on EuroSpeech demonstrates sizable improvements on out-of-domain benchmarks, illustrating practical benefits for low-resource languages. The work provides public data and tooling to lower barriers to multilingual speech research and dataset creation.

Abstract

Recent progress in speech processing has highlighted that high-quality performance across languages requires substantial training data for each individual language. While existing multilingual datasets cover many languages, they often contain insufficient data for most languages. Thus, trained models perform poorly on the majority of the supported languages. Our work addresses this challenge by introducing a scalable pipeline for constructing speech datasets from parliamentary recordings. The proposed pipeline includes robust components for media retrieval and a two-stage alignment algorithm designed to handle non-verbatim transcripts and long-form audio. Applying this pipeline to recordings from 22 European parliaments, we extract over 61k hours of aligned speech segments, achieving substantial per-language coverage with 19 languages exceeding 1k hours and 22 languages exceeding 500 hours of high-quality speech data. We obtain an average 41.8\% reduction in word error rates over baselines when finetuning an existing ASR model on our dataset, demonstrating the usefulness of our approach.

Figures (4)

• Figure 1: Overview of the EuroSpeech data processing pipeline. The workflow begins with the initial Data Sourcing and Metadata Collection phase, which gathers metadata from parliamentary websites and APIs. This structured information (as Links CSV) feeds into the Download Pipeline to retrieve raw audio and transcripts. The Raw Audio and Transcripts are then processed by the Alignment Pipeline, which segments the audio and matches it to the corresponding text. The final output is the Aligned Dataset, consisting of short audio segments paired with their transcriptions, ready for model training.
• Figure 2: The broad European geographic coverage of EuroSpeech. Countries are colored by the total hours of aligned speech data (CER < 30%) available in the dataset. A perceptually uniform color scale is used.
• Figure 3: Audio duration for each language in EuroSpeech after the alignment pipeline and at different Character Error Rate (CER) filtering stages (CER < 30%, < 20%, and < 10%). Languages are ordered by their data volume at CER < 20%. A dashed horizontal lines indicates the 1,000-hour and 500-hour thresholds. EuroSpeech showcases large amounts of low-CER data across its languages.
• Figure 4: Language distribution in the EuroSpeech CER < 30% subset, highlighting a key strength of our dataset: the balanced distribution across multiple languages rather than concentration in just a few dominant ones.