Finnish speech models improve using confidence based filtering and model comparisons
From 2.1M Hours of Raw Broadcast Audio to Finnish Speech SSL: Confidence Filtering and Architectural Analysis at Scale
Computation and LanguageSound
Summary
Working with millions of hours of Finnish broadcast audio, the authors studied how to pick good speech segments for teaching AI to understand spoken language better. Instead of just checking if speech is present, they used a tool that guesses how confident it is about the speech content to filter out poor-quality data. This approach led to clearer AI understanding with fewer errors. They also compared different computer designs for making these speech models, finding some worked better for Finnish speech.
What this means in practice
- •For speech technology developers: Produce better Finnish speech recognition models by filtering training data using ASR confidence scores instead of standard language ID methods.
- •For ai infrastructure architects: Choose among Transformer, Macaron-Net, Conformer, and E-Branchformer architectures with confidence for Finnish speech tasks based on matched model size comparisons in this study.
Authors
Yaroslav Getman, Tamás Grósz, Tommi Lehtonen, Mikko Kurimo
Abstract
Broadcast archives offer scale and diversity for self-supervised learning (SSL) but contain substantial noisy, non-speech, and multilingual content. Prior work typically relies on voice activity detection or language identification (LID) to extract speech, ensuring presence rather than quality. Whether and how to filter unlabeled speech by quality for SSL, particularly from noisy broadcasts, remains underexplored. We present a large-scale Finnish SSL study starting from 2.1M hours of unconstrained broadcast audio. We propose automatic speech recognition (ASR) confidence as a proxy for pretraining data quality and show 3-7% relative WER improvements over LID-based filtering despite a smaller training set. We further address the lack of controlled architectural comparisons in speech SSL by evaluating Transformer, Macaron-Net, Conformer, and E-Branchformer at matched parameter budgets. Our models improve upon prior Finnish SSL baselines, particularly at Base scale and out-of-domain.