Nord-Parl-TTS: Finnish and Swedish TTS Dataset from Parliament Speech

1. INTRODUCTION

• Low-resource languages lack the public recordings and speaker access needed for conventional high-quality TTS datasets, slowing and raising the cost of system development.Recent large-scale in-the-wild efforts mainly target high-resource languages, leaving lower-resource languages overlooked.Source: §1 · p. 1
• The dataset adapts Emilia’s pipeline to parliamentary recordings, yielding 900 hours of Finnish and 5090 hours of Swedish TTS data.Parliament recordings are open-access and may be processed and distributed if their original content is not modified.Source: §1 · p. 1
• The release includes unified evaluation sets for Finnish and Swedish TTS.These sets are intended to support development and benchmarking.Source: §1 · p. 1
• The authors benchmark two representative open-source TTS systems to demonstrate the dataset and evaluation sets’ applicability to model development and evaluation.Source: §1 · p. 1

2. RELATED WORK

• Finnish TTS resources include Perso Synteesi (20 hours), CSS10 (10 hours), and FinSyn (approximately 60 hours); the authors know of no existing public Swedish TTS dataset.Perso Synteesi and FinSyn are also not conveniently available from open-sourced data hubs.Source: §2 · pp. 1–2
• The Finnish Parliament ASR corpus is not an optimal TTS source because transcript corrections alter spontaneous speech and word-level segmentation can disrupt sentence-level prosody modeling.The transcripts replace hesitations, repetitions, and slips with written-language equivalents, while utterances may begin or end mid-sentence.Source: §2 · p. 2
• RixVox-v2 provides nearly 23000 hours of Swedish parliamentary ASR speech, standardized to 16 kHz mono and segmented into utterances up to 30 seconds.Source: §2 · p. 2

3.1. Finnish

• The Finnish pipeline standardizes extracted audio to 24 kHz mono, separates noise, diarizes speakers, applies fine-grained VAD, and excludes samples with DNSMOS P.835 OVRL below 3.0.If normalized-transcript WER between Whisper and Finnish-only Wav2Vec2-large is below 5%, Whisper’s transcript is considered a candidate, not automatically retained; Whisper is chosen for capitalization and punctuation.Source: §3.1 · p. 2
• Finnish evaluation uses 500 Perso Synteesi prompt-target pairs, balanced between male and female speakers and restricted to 3–20 seconds and content longer than 10 characters.Source: §3.1 · pp. 2–3

3.2. Swedish

• The Swedish pipeline resamples RixVox-v2 audio to 24 kHz, removes non-speech and multi-speaker segments, filters WER above 10% and DNSMOS below 3.0, yielding 5090 hours.The 10% WER threshold accommodates transcript corrections while aiming to retain scale; the resulting duration distribution increases monotonically.Source: §3.2 · p. 3
• Swedish evaluation candidates come from Common Voice Swedish 22.0, are screened by duration, gender labels, speaker count, and human listening, then formed into 500 prompt-target pairs.The listening screen removes unclear or quiet speech, strong non-native accents, and background noise.Source: §3.2 · p. 3

4. EXPERIMENTS

• Matcha-TTS uses monotonic alignment search, while F5-TTS learns implicit alignment through Diffusion Transformers; both are diffusion-based non-autoregressive systems.No autoregressive system is included because the authors could not find official open-source implementations for typical AR TTS models.Source: §4 · p. 3
• Matcha-TTS is trained for 500k updates on one A100 per language, while F5-TTS follows its v1 Base configuration with 24 MI250X GPUs and 1.2M updates.Matcha uses batch size 64; F5-TTS uses a global batch of 308400 mel-spectrogram frames.Source: §4 · p. 3
• Objective evaluation measures CER against target content and cosine speaker similarity to the prompt; subjective evaluation uses CMOS for human-likeness and SMOS for speaker similarity.Submissions failing language skill checks are excluded from the subjective results.Source: §4 · p. 4

• CER > 100% filtering excludes 4 Finnish and 18 Swedish F5-TTS-Base utterances, versus one utterance in each language for Matcha-TTS.Source: §5 · p. 4
• Matcha-TTS records lower CER than F5-TTS-Base in Finnish (2.55% vs 6.72%) and Swedish (4.66% vs 13.64%).Source: Table 2 · p. 4, §5 · p. 4
• Explicit alignment modeling is reported to reduce hallucination and improve intelligibility despite Matcha-TTS having fewer parameters, a smaller batch size, and fewer updates.Source: §5 · p. 4
• Matcha-TTS has higher Finnish SIM than F5-TTS-Base (0.566 vs 0.538), while F5-TTS-Base has higher Swedish SIM (0.53 vs Matcha-TTS’s 0.442).Source: §5 · p. 4
• Subjectively, Matcha-TTS scores Finnish CMOS -1.93 and SMOS 2.49, and Swedish CMOS -2.25 and SMOS 2.27; Swedish F5-TTS-Base scores -1.46 and 3.41, respectively.Source: §5 · p. 4

• Open-sourced TTS models trained on Nord-Parl-TTS achieve acceptable synthesis results, supporting the dataset’s applicability for TTS training.Source: §6 · p. 4
• Future work includes refining tooling, releasing curation tools and visualizations, adding Danish and Norwegian, providing more benchmark models, and improving dataset quality.Source: §6 · p. 4