💬 NLP

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

This paper introduces the Swiss Parliaments Corpus Re-Imagined (SPC_R), a high-quality long-form dataset of 555 hours of Swiss German parliamentary debates created by transcribing audio with Whisper Large-v3, refining it through a two-step GPT-4o correction process, and filtering segments based on predicted BLEU and semantic completeness scores to achieve a significant 6-point BLEU improvement over previous sentence-level releases.

Original authors: Vincenzo Timmel, Manfred Vogel, Daniel Perruchoud, Reza Kakooee

Published 2026-03-13

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Vincenzo Timmel, Manfred Vogel, Daniel Perruchoud, Reza Kakooee

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you have a massive, dusty library of old radio broadcasts from the Swiss Parliament. These aren't just any broadcasts; they are hours-long debates spoken in Swiss German (a tricky, dialect-heavy version of the language), but the official written records are in Standard German.

The goal of this paper is to build a perfect "dictionary" that matches the spoken audio to the written text, so computers can learn to understand Swiss German. The authors call this new library SPC_R.

Here is how they built it, explained through a simple story:

1. The Problem: The "Hasty Scribe"

Imagine you hire a very fast, very smart robot scribe (called Whisper) to listen to these 801 hours of debates and type them out.

The Good News: The robot is fast and gets the general idea right.
The Bad News: It's a bit careless with details. It hears "Alba Rutschi" but writes "Alberucci." It hears a specific political party name and gets it wrong. It's like a student who understands the story but keeps misspelling the main characters' names.

2. The Solution: The "Two-Step Editor"

To fix the robot's mistakes, the authors didn't just hire a human to read every single word (that would take forever). Instead, they built a smart editing pipeline using two different AI "editors."

Step A: The Fact-Checker (RAG + GPT-4o)

First, they gave the robot's messy draft to a super-smart editor (GPT-4o). But this editor had a secret weapon: Context.

The Analogy: Imagine the editor is trying to fix a sentence about a politician named "Hans." Without context, the editor might guess. But, the system also feeds the editor the official meeting minutes (the source of truth) right next to the audio.
The Magic: Using a technique called RAG (Retrieval-Augmented Generation), the system finds the exact paragraph in the official minutes that matches the audio snippet. The editor then says, "Ah, the robot wrote 'Alberucci', but the official minutes say 'Alberucci'. Let's fix it."
Result: This step fixed almost all the name errors, turning a 72% accuracy rate into 100% for names and places.

Step B: The Quality Inspector (GPT-4o-mini)

Once the text was cleaned up, they needed to know: "Is this segment actually good enough to keep?"

The Analogy: Imagine a film director watching a raw take. Some scenes are perfect; others are mumbled or confusing. The director needs to decide which clips to keep for the final movie.
The Tool: They used a second, smaller AI (GPT-4o-mini) to act as the "Judge." It looked at the corrected text and gave it a score from 0 to 3.
- Score 3: Perfect.
- Score 2: Minor grammar slip, but the meaning is clear.
- Score 1: Wrong names or key words.
- Score 0: Garbage.

3. The "Crystal Ball" (Predicting Quality without Humans)

Here is the cleverest part. Usually, to know if a transcription is good, you need a human to compare it to the "real" answer. That's expensive and slow.

The authors discovered a crystal ball inside the robot scribe (Whisper).

The Analogy: When the robot scribes a word, it whispers to itself, "I'm 99% sure this is 'cat'." If it says "I'm only 50% sure," that's a red flag.
The Discovery: They found a direct link between how "confident" the robot was and how good the final text was. They built a math formula (a "Predicted BLEU") that looks at the robot's confidence and predicts the quality score before a human ever looks at it.
The Result: They could automatically filter out the bad clips without hiring a single human to read them first.

4. The Final Product

By combining these steps, they started with 801 hours of raw audio.

They filtered out the low-confidence, messy parts.
They kept only the high-quality, corrected segments.
The Result: A pristine library of 555 hours of Swiss German audio perfectly matched with Standard German text.

Why Does This Matter?

Think of this dataset as training wheels for AI.
Before this, AI models trying to understand Swiss German were like students trying to learn a language with a textbook full of typos. Now, they have a textbook that has been proofread by a super-intelligent editor and verified by a strict judge.

This new dataset allows AI to:

Understand Swiss German much better (improving accuracy by 6 points, which is huge in this field).
Get the names of people and places right (crucial for real-world use).
Learn from long, continuous conversations rather than just short, isolated sentences.

In short, the authors took a messy, difficult-to-use pile of data, ran it through a high-tech factory of AI editors and filters, and shipped out a premium product that will help computers finally "speak" Swiss German.

1. Problem Statement

The development of Automatic Speech Recognition (ASR) systems for Swiss German is hindered by data scarcity, particularly in low-resource, domain-specific contexts.

Limitations of Existing Data: Previous datasets, such as the original Swiss Parliaments Corpus (SPC) by Plüss et al. (2021), are segmented at the sentence level. This segmentation fails to capture the context necessary for real-world ASR applications, which often require long-form processing.
Quality Issues: Raw transcriptions from standard ASR models often contain significant errors in named entities (e.g., names, places, political parties) and dialect-specific nuances, which degrade the utility of the data for training robust models.
Need for High-Quality Long-Form Data: There is a critical need for large-scale, long-form audio-text pairs that align Swiss German speech with Standard German transcripts, processed with high accuracy to support advanced NLP tasks.

2. Methodology

The authors propose a sophisticated pipeline to convert 801 hours of raw Swiss German parliamentary debate audio into a high-quality corpus. The process involves three main stages:

A. Initial Transcription (Whisper Large-v3)

Model: The pipeline utilizes Whisper Large-v3 under high-compute settings (beam_size=10, best_of=10, log_prob_threshold=-2) on NVIDIA A4500 GPUs.
Output: This generates initial transcriptions and captures the model's avg_log_prob (average token log-probability), which serves as a confidence metric.

B. Quality Prediction (Predicted BLEU)

Concept: The authors establish a linear relationship between Whisper's confidence metric and the actual BLEU score.
Mechanism: Using linear regression on a small set of manually verified ground truths, they derived a formula to predict the BLEU score solely based on Whisper's avg_log_prob.
- Equation: $Predicted\ BLEU = 1.59 \times Confidence - 0.68$
Application: This allows the authors to filter out low-quality segments without needing a human reference for every single segment, enabling scalable quality control.

C. Two-Step LLM Correction & Evaluation

To address errors in named entities and dialect nuances, a two-step Large Language Model (LLM) process is applied:

Correction Stage (RAG-based):
- Retrieval-Augmented Generation (RAG): The system retrieves relevant context from official manual session protocols (summaries) using a vector store (Faiss) and text-embedding-3-large.
- Correction: GPT-4o is prompted to correct the raw Whisper transcription segment-by-segment, using the retrieved official protocol as ground-truth context. This specifically targets named entities and dialect-specific corrections (e.g., correcting "vo dr" to "von der").
Evaluation Stage:
- Judgment: A separate model, GPT-4o-mini, acts as a "Judge" to evaluate the corrected segments.
- Scoring: Segments are categorized on a scale (0–3) based on semantic completeness and accuracy of key terms (names, numbers, abbreviations).
- Filtering: Segments scoring below a certain threshold (specifically merging categories 0/1 and 2/3 for reliability) are discarded.

3. Key Contributions

SPC_R Corpus Release: A new dataset containing 555 hours of high-quality, long-form Swiss German audio paired with Standard German text (derived from an initial 801 hours).
Predicted BLEU Metric: A novel method to estimate transcription quality (BLEU score) directly from ASR confidence scores, facilitating automated data filtering.
RAG-Enhanced Correction Pipeline: A demonstration of using RAG with LLMs to significantly improve named entity recognition and dialect handling in ASR post-processing.
Long-Form Focus: Unlike previous sentence-level datasets, SPC_R preserves the natural flow and context of multi-hour parliamentary debates.

4. Results

Dataset Size: The final corpus comprises 555 hours of validated audio (filtered from 801 hours).
- Train/Test Split: 525 hours for training, 30 hours for testing (with stricter quality criteria: BLEU $\ge$ 70 and Category 3 evaluation).
Quality Improvements:
- Word Error Rate (WER): Reduced from 15.7% (Whisper Large-v3 standard) to 11.1% after GPT-4o correction.
- Named Entity Accuracy: Improved from 72.2% (52/72 correct) to 100% (72/72 correct) for names, places, and legal references.
- BLEU Score: The long-form dataset achieved a 6-point BLEU improvement compared to the original sentence-level SPC release.
Filtering Efficiency: The Predicted BLEU threshold of 65 effectively filtered out lower-quality segments while retaining a substantial volume of data.

5. Significance and Impact

Advancing Low-Resource NLP: SPC_R provides a critical resource for training robust ASR models for Swiss German, a dialect often underrepresented in high-quality datasets.
Methodological Blueprint: The paper demonstrates a scalable workflow for creating high-quality speech corpora by combining:
1. High-compute ASR.
2. Predictive quality filtering (Predicted BLEU).
3. Context-aware LLM correction (RAG).
4. Automated LLM evaluation.
Commercial and Research Utility: The dataset is released under CC BY 4.0, making it accessible for both research and commercial applications, addressing a gap left by previous datasets restricted to research-only use.
Future Directions: The authors suggest expanding data sources beyond Bernese debates to increase dialectal diversity and exploring open-source LLMs to reduce costs.

Conclusion

The SPC_R project successfully transforms raw, noisy parliamentary audio into a curated, high-fidelity speech-text corpus. By leveraging the synergy between robust ASR models, retrieval-augmented LLM correction, and data-driven filtering, the authors have created a dataset that significantly outperforms previous iterations in both volume and quality, setting a new standard for Swiss German speech processing.