Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Swiss Parliaments Corpus Re-Imagined (SPC_R): Enhanced Transcription with RAG-based Correction and Predicted BLEU

Vincenzo Timmel, Manfred Vogel, Daniel Perruchoud, Reza Kakooee

Abstract

This paper presents a new long-form release of the Swiss Parliaments Corpus, converting entire multi-hour Swiss German debate sessions (each aligned with the official session protocols) into high-quality speech-text pairs. Our pipeline starts by transcribing all session audio into Standard German using Whisper Large-v3 under high-compute settings. We then apply a two-step GPT-4o correction process: first, GPT-4o ingests the raw Whisper output alongside the official protocols to refine misrecognitions, mainly named entities. Second, a separate GPT-4o pass evaluates each refined segment for semantic completeness. We filter out any segments whose Predicted BLEU score (derived from Whisper's average token log-probability) and GPT-4o evaluation score fall below a certain threshold. The final corpus contains 801 hours of audio, of which 555 hours pass our quality control. Compared to the original sentence-level SPC release, our long-form dataset achieves a 6-point BLEU improvement, demonstrating the power of combining robust ASR, LLM-based correction, and data-driven filtering for low-resource, domain-specific speech corpora.

Swiss Parliaments Corpus Re-Imagined (SPC_R): Enhanced Transcription with RAG-based Correction and Predicted BLEU

Abstract

This paper presents a new long-form release of the Swiss Parliaments Corpus, converting entire multi-hour Swiss German debate sessions (each aligned with the official session protocols) into high-quality speech-text pairs. Our pipeline starts by transcribing all session audio into Standard German using Whisper Large-v3 under high-compute settings. We then apply a two-step GPT-4o correction process: first, GPT-4o ingests the raw Whisper output alongside the official protocols to refine misrecognitions, mainly named entities. Second, a separate GPT-4o pass evaluates each refined segment for semantic completeness. We filter out any segments whose Predicted BLEU score (derived from Whisper's average token log-probability) and GPT-4o evaluation score fall below a certain threshold. The final corpus contains 801 hours of audio, of which 555 hours pass our quality control. Compared to the original sentence-level SPC release, our long-form dataset achieves a 6-point BLEU improvement, demonstrating the power of combining robust ASR, LLM-based correction, and data-driven filtering for low-resource, domain-specific speech corpora.

Paper Structure

This paper contains 13 sections, 1 equation, 7 figures, 1 table.

Table of Contents

  1. Introduction
  2. Related Work
  3. Transcription with Whisper Large-v3
  4. BLEU Prediction
  5. Transcript correction using GPT-4o
  6. Context provision via RAG
  7. Correction Stage
  8. Evaluation Stage
  9. Selecting Data and Train/Test Split
  10. Availability and License
  11. Conclusion
  12. Future Work
  13. Limitations

Figures (7)

  • Figure 1: Overview of Swiss German speech to German text datasets. Usage of SPC is possible under MIT license, SDS-200 and STT4SG-350 under SwissNLP license. SwissDial can be used exclusively for research purposes.
  • Figure 2: Linear relationship between BLEU score vs. Whisper confidence score for ten long-form conversations, represented by numbers 1-10. The blue shaded area represents the 95% confidence interval.
  • Figure 3: Distribution of Predicted BLEU scores across SPC_R ($N$ = 131'291 data segments).
  • Figure 4: Percentage of data samples that have a BLEU score above the threshold.
  • Figure 5: Word Error Rates (WER) for Whisper Large‑v3 under three configurations: standard settings, after applying GPT‑4o correction, and using high-compute settings (enhanced settings) with GPT‑4o correction.
  • ...and 2 more figures