Source-linked AI summary
WenetSpeech-Min: A Large-Scale Minnan Speech Corpus with Dual Transcriptions for Dialectal Speech Processing
Haoyu Zhang, Chunjiang He, Hongtao Li, Zeyu Zhu, Qituan Shangguan, Chengyou Wang, Jingbin Hu, Ziyu Zhang, Bingshen Mu, Yanbo Wang, Shuai Wang, Jinhui Ye, Chengdong Liang, Binbin Zhang, Pengcheng Zhu, Chuang Ding, Qianze Feng, Qingyang Hong, Liumeng Xue, Lei Xie
TL;DR
Minnan speech technology lacks large-scale real-world corpora with paired Minnan and Mandarin transcripts, as well as reproducible benchmarks. WenetSpeech-Min provides a roughly 10,000-hour paired-transcript corpus and ASR and TTS benchmarks; its trained models outperform evaluated open-source systems on most metrics and are competitive with commercial systems.
Problem
Minnan has limited large-scale public speech corpora, and large-scale real-world speech with paired Minnan and Mandarin transcripts and reproducible benchmarks remains unavailable.
Source: §1 · p. 1Method
The paper builds a multi-source corpus of approximately 10,000 hours with paired Minnan and Mandarin transcripts, establishes ASR and TTS benchmarks, and trains models on the corpus.
Source: §1 · p. 1Results
Models trained on WenetSpeech-Min outperform evaluated open-source baselines on most metrics and are competitive with commercial systems.
Source: §5 · p. 4Takeaways & Limitations
The corpus and benchmarks address the lack of comprehensive evaluation resources for Minnan speech technology.
Source: §5 · p. 4
Abstract
from arXiv · showhide
Progress in dialectal speech technology is hindered by the scarcity of large-scale, real-world corpora. For Minnan speech, existing resources remain limited, and few provide paired Minnan and Mandarin transcripts at scale. To address these gaps, we introduce WenetSpeech-Min, an open-source corpus comprising around 10,000 hours of Minnan speech collected from diverse online media, with paired Minnan and Mandarin transcripts for every utterance. We further establish an automatic speech recognition (ASR) benchmark covering both Minnan and Mandarin transcripts and a text-to-speech synthesis (TTS) benchmark using Minnan transcripts, with manually verified evaluation sets for both tasks. To assess the effectiveness of the corpus, we train ASR and TTS models on WenetSpeech-Min and compare them with representative systems on the proposed benchmarks. The resulting models outperform the evaluated open-source models on most metrics and achieve competitive performance against commercial systems. We will release the corpus, benchmarks, and models to facilitate reproducible research on Minnan speech technology.
1. INTRODUCTION
Minnan remains under-resourced despite being widely spoken, and large-scale real-world speech with paired Minnan and Mandarin transcripts and reproducible benchmarks is lacking. WenetSpeech-Min addresses these gaps with a large paired-transcript corpus, public ASR and TTS benchmarks, and trained models that perform strongly against evaluated systems.
1. INTRODUCTION
- Minnan is widely spoken across Fujian, Taiwan, and Southeast Asia but remains under-resourced in speech technology, with limited large-scale public corpora.Source: §1 · p. 1
- Large-scale real-world Minnan speech with paired dialect and Mandarin transcripts remains unavailable, and public benchmarks for paired-transcript ASR and Minnan TTS are lacking.Existing resources provide only small read speech, Mandarin-only transcripts, or limited paired-transcript data; some evaluation data are not public or do not support dialect-transcript ASR.Source: §1 · p. 1
- WenetSpeech-Min is a public, approximately 10,000-hour multi-source corpus of real-world Minnan speech, pairing every utterance with Minnan and Mandarin transcripts.The paper also establishes paired-transcript ASR and Minnan TTS benchmarks with manually verified evaluation sets and unified evaluation protocols.Source: §1 · p. 1
- Models trained on WenetSpeech-Min outperform evaluated open-source models on most metrics and remain competitive with commercial systems.Source: §1 · p. 1
2. WENETSPEECH-MIN
WenetSpeech-Min combines a large paired-transcript Minnan dataset with a pipeline for processing heterogeneous media and generating transcripts. Its released material includes relatively clean speech and challenging legacy recordings, alongside predominantly short utterances and some longer segments.
2.1. Data Construction Pipeline
- The construction pipeline has two components—audio processing and paired-transcript generation—overviewed in Fig. 1.Audio segments are filtered by speech content and acoustic quality before transcript generation.Source: §2.1 · pp. 1–2, Figure 1 · p. 2
2.1.1. Audio Processing
- The audio-processing stage segments online recordings with VAD, estimates SNR, duration, and WV-MOS, filters low speech-confidence segments, and stores speaker-count metadata.The DaSheng audio-event-detection model estimates speech confidence, and pyannote estimates the number of speakers.Source: §2.1.1 · p. 2
2.1.2. Annotation Tool Construction
- Manually verified Minnan transcripts from 1,600 hours fine-tune three ASR annotators, while a paired 600-hour subset trains a Minnan-to-Mandarin translation model.The translation model is based on Qwen3-8B fine-tuned with LoRA.Source: §2.1.2 · p. 2
2.1.3. Paired-Transcript Generation
- Three ASR annotators produce complementary dialect hypotheses that are aligned and combined with ROVER to form each dialect transcript.Source: §2.1.3 · p. 2
- Mandarin transcripts come from OCR-extracted embedded subtitles or translation of the dialect transcript, followed by normalization and consistency checks.Pairs with unreliable segmentation or low-confidence content are removed or returned for correction.Source: §2.1.3 · p. 2
2.2.1. Data Scale and Source Distribution
- Approximately 10,000 hours form 5,121,249 utterances, each carrying dialect and Mandarin transcripts; drama accounts for 54.7% of sourced hours.The domain distribution covers roughly 8,500 hours with recorded provenance; Audiobook contributes 19.8% and Entertainment 13.9%.Source: §2.2.1 · p. 2
2.2.2. Audio Quality
- Archival television averages WV-MOS 1.09 versus 2.47 for general-source speech, while 73.8% of released duration falls between 10 and 40 dB SNR.The lower archival scores largely reflect inherited recording, broadcast, and archival conditions; the retained material adds natural multi-speaker interaction, emotional speech, and varied speaking styles.Source: §2.2.2 · p. 2, Figure 2 · p. 3
- Most utterances are short, with a secondary peak around 20 seconds from pre-segmented source clips retained without another VAD pass.These longer contiguous segments provide complementary training examples with longer temporal context.Source: §2.2.2 · pp. 2–3
3. BENCHMARK
The benchmarks evaluate paired-transcript ASR on shared audio and Minnan TTS with expert-verified text and reference speakers. Both use defined objective measures, with the TTS benchmark also including native-listener ratings.
3.1. Paired-Transcript ASR Benchmark
- WS-Min-Eval-ASR pairs six hours of audio with manually verified Minnan and Mandarin references, scored using D-CER for Minnan and M-BLEU and M-CER for Mandarin.All systems use a unified text-normalization and scoring procedure; M-BLEU complements character-level accuracy when Mandarin renderings differ in wording.Source: §3.1 · p. 3
3.2. Minnan TTS Benchmark
- WS-Min-Eval-TTS has 500 Easy and 250 Hard texts, nine reference speakers, objective CER and SIM measures, and native-listener ratings on 40 test cases.The Hard subset includes long-form texts, repetition-heavy texts, and tongue twisters; listeners rate intelligibility, speaker similarity, and accent authenticity.Source: §3.2 · p. 3
4. EXPERIMENTS
The experiments evaluate ASR for both Minnan and Mandarin transcription and assess Minnan TTS against representative systems. WenetSpeech-Min-trained models perform strongly across these tasks, with benchmark-specific differences in results.
4.1. Automatic Speech Recognition
- ASR systems are evaluated under their supported output settings, while the Mandarin models include MinSpeech and WenetSpeech-Min training variants.For Minnan transcription, Qwen3-ASR is fine-tuned on WenetSpeech-Min; FireRedASR2-AED is also fine-tuned on paired transcripts to support both outputs.Source: §4.1 · p. 3
- 15.21% D-CER establishes a new state of the art for Minnan-transcript recognition on WS-Min-Eval-ASR after further internal-data training.Qwen3-ASR-WSM-Min outperforms evaluated existing systems for Minnan transcription; the cited table reports Minnan results on WS-Min-Eval-ASR.Source: §4.1 · p. 3, Table 2 · p. 3
- 33.05% average M-CER and 54.24 average M-BLEU give the internally trained Mandarin model the lowest average M-CER and second-highest average M-BLEU.The M-BLEU result nearly matches the best commercial result of 54.25; averages are unweighted across three test sets.Source: §4.1 · p. 3, Table 3 · p. 4
- FireRedASR2-AED-WSM achieves the best GigaSpeechBench results while remaining competitive on Minnan and Mandarin transcription.Paired-transcription training supports both outputs within one model; absolute M-CER and M-BLEU scores are not directly comparable across benchmarks.Source: §4.1 · p. 4
4.2. Speech Synthesis
- CosyVoice3-WSM is built through continual pre-training, filtered supervised fine-tuning, and preference optimization.The supervised stage uses 1,000 hours selected by SNR and MOS; preference pairs are filtered using CER and SIM.Source: §4.2 · p. 4
- CosyVoice3-WSM lowers CER and improves SIM on both TTS subsets, and is comparable to commercial systems overall.Among evaluated open-source systems, it has the lowest CER and highest I-MOS and S-MOS on both subsets, plus the highest A-MOS on Hard; its SIM remains below FireRedTTS3 and VoxCPM2.Source: §4.2 · p. 4, Table 4 · p. 4
5. CONCLUSION
The paper introduces a large, multi-source Minnan speech corpus with paired Minnan and Mandarin transcripts, alongside ASR and TTS benchmarks. Models trained on the corpus outperform evaluated open-source baselines on most metrics and are competitive with commercial systems.
5. CONCLUSION
- WenetSpeech-Min contains approximately 10,000 hours of multi-source Minnan audio paired with Minnan and Mandarin transcripts.The corpus is presented as a large-scale resource.Source: §5 · p. 4
- The paper establishes paired-transcript ASR and Minnan TTS benchmarks to address the lack of comprehensive evaluation resources.Source: §5 · p. 4
- Models trained on WenetSpeech-Min outperform evaluated open-source baselines on most metrics and are competitive with commercial systems.Source: §5 · p. 4
7. COMPLIANCE WITH ETHICAL STANDARDS
The study uses publicly accessible online speech data solely for research and reports consent, anonymization, and institutional and ethical compliance.
7. COMPLIANCE WITH ETHICAL STANDARDS
- Speech data came from publicly accessible online media and were used solely for research purposes.Source: §7 · p. 5
- Reference speakers and listening-test participants provided informed consent, and their responses were anonymized.Source: §7 · p. 5
- The study was conducted in accordance with applicable institutional and ethical requirements.Source: §7 · p. 5