WenetSpeech-Chuan: Sichuanese Dialect Corpus

• WenetSpeech-Chuan is a 10,013-hour annotated corpus distinguishing strong and weak labels to support robust automatic speech recognition and text-to-speech training.
• It utilizes the Chuan-Pipeline, which integrates preprocessing, speaker diarization, LLM-assisted transcript correction, and multimodal punctuation to ensure dialect authenticity.
• Benchmark results demonstrate significant improvements in CER and accent naturalness, addressing data scarcity and AI equity issues in dialect speech technology.

WenetSpeech-Chuan is a 10,000-hour, richly annotated open-source corpus for Sichuanese dialect speech, accompanied by a complete data processing framework, “Chuan-Pipeline,” and manually verified evaluation benchmarks for automatic speech recognition (ASR) and text-to-speech (TTS). It is positioned as the largest open-source corpus for Sichuanese dialects and is intended to address the chronic lack of large-scale dialect resources for speech technology, especially for a dialect group spoken by roughly 120 million people. The release includes the corpus itself, WenetSpeech-Chuan-Eval benchmarks, trained models, and processing recipes on a public project page (Dai et al., 22 Sep 2025).

1. Corpus scope and composition

WenetSpeech-Chuan is organized by transcription confidence rather than by a single homogeneous training split. The corpus contains 10,013 hours of raw audio, partitioned into Strong Label and Weak Label subsets according to ASR confidence. Strong-label data is intended for fully supervised training, whereas weak-label data is preserved for semi-supervised or self-supervised use (Dai et al., 22 Sep 2025).

Set	Confidence range	Duration (h)
Strong Label	[0.9, 1.0]	3,714
Weak Label	[0.6, 0.9)	6,299
Total	/	10,013

The corpus targets Sichuan-Chongqing dialects within the broader Sichuanese group. The design premise is that these dialects differ from Standard Mandarin in the tonal system, vocabulary, and grammar, so the resource is framed as a dialect corpus rather than as a simple accented-Mandarin extension (Dai et al., 22 Sep 2025). This distinguishes it from much smaller resources such as MagicData-Conversation and MagicData-Daily-Use, and from corpora such as KeSpeech that include Southwestern accents but are characterized here as accented Mandarin rather than a true dialectal resource.

The data is drawn from highly heterogeneous web sources. The domain distribution is dominated by short videos (52.83%), entertainment (20.08%), and live streams (18.35%), with smaller but important coverage of documentaries, audiobooks, interviews, news, reading, and drama. Long recordings are segmented by VAD into 5–25 second clips, which preserves natural spontaneity while making the data tractable for large-scale ASR and TTS training (Dai et al., 22 Sep 2025).

Audio quality is not “studio-clean” by construction. Instead, the corpus intentionally preserves realistic acoustic variability from user-generated and broadcast-like media. Quality control is enforced through Word-level Virtual MOS (WVMOS) based on SNR and duration features, and the resulting quality distribution peaks around WVMOS 3.0–3.5, indicating mostly moderate-to-high quality rather than narrow, curated laboratory conditions (Dai et al., 22 Sep 2025).

2. Annotation schema and metadata

A central feature of WenetSpeech-Chuan is that it is richly annotated beyond transcription alone. Each utterance carries multiple labels spanning speaker identity, demographic or paralinguistic attributes, emotion, and domain. This makes the corpus suitable not only for ASR and TTS, but also for controlled analysis of dialect robustness across speaker types and speaking conditions (Dai et al., 22 Sep 2025).

Speaker identity is obtained by first isolating single-speaker segments with pyannote diarization and then extracting speaker embeddings with CAM++. Clustering assigns a consistent speaker ID to utterances attributed to the same individual. The paper does not report a total speaker count, but speaker tracking is a first-class annotation layer in the release.

Paralinguistic labels include gender, age, and emotion. Gender is assigned by a pre-trained classifier with a reported 98.7% accuracy. Age is discretized into five stages: children, teenager, young, middle-aged, and old. Emotion is inferred by combining Emotion2vec and SenseVoice predictions through majority voting over seven classes: happy, angry, sad, neutral, fearful, surprised, and disgusted (Dai et al., 22 Sep 2025).

Domain labels are also retained. The same nine-domain ontology used in corpus construction is carried into evaluation design, which allows systematic contrast between clean, read speech and noisy, spontaneous media speech. A practical consequence is that evaluation can be stratified into “Easy” and “Hard” subsets without collapsing all dialectal performance into a single average.

Transcriptions are automatically generated and then punctuation is added through a multimodal process. Punctuation is not treated as a purely textual post-processing problem. Instead, Kaldi force alignment provides word timestamps and pause durations, while a BiLSTM punctuation model predicts punctuation types. The system treats pauses of at least 0.25 s as candidate short pauses corresponding to commas and pauses of at least 0.5 s as candidate long pauses corresponding to periods, question marks, or exclamation marks. These thresholds are iteratively refined with human feedback so that punctuation tracks real prosodic boundaries rather than only text-based syntax (Dai et al., 22 Sep 2025).

3. Chuan-Pipeline

Chuan-Pipeline is the processing framework used to transform raw web audio into the released corpus. It combines source selection, segmentation, diarization, quality assessment, LLM-assisted transcription correction, and multimodal punctuation prediction into a single dialect-oriented workflow (Dai et al., 22 Sep 2025).

The first stage is preprocessing and labeling. Metadata is mined from online video and streaming platforms, and candidate content is filtered to identify material likely to contain Sichuanese. The paper states that candidate streams then undergo manual verification to confirm dialect presence. After this, VAD segments long recordings into 5–25 second clips, and pyannote diarization is used to isolate single-speaker segments. Speaker embeddings from CAM++ support clustering and speaker-ID assignment.

The second stage is quality assessment. Chuan-Pipeline computes SNR- and duration-based features and combines them into WVMOS. Segments with low WVMOS are discarded. This stage is important because the source domains include short-video and live-stream conditions where background noise, compression artifacts, and music contamination can be substantial.

The third stage is LLM-GER, described as Generative Error Correction based ROVER. Each clip is decoded by three ASR systems: FireRed-ASR, SenseVoice-small, and TeleASR. Their outputs are then fed to Qwen3, which is prompted to exploit its Sichuanese dialectal knowledge while preserving semantics and token length. The resulting corrected transcript serves as a fourth hypothesis, and the system computes a confidence score from the four transcriptions to decide whether a sample belongs in the strong-label or weak-label subset. On held-out data, the paper reports that LLM-GER yields about 15% transcription accuracy improvement over individual systems (Dai et al., 22 Sep 2025).

The fourth stage is multimodal punctuation prediction. Audio-side force alignment yields word timestamps and pause intervals, and a text-side BiLSTM predicts punctuation class. The combined system produces punctuated transcripts that are better aligned with speaking rhythm and thus more useful for downstream TTS than plain unpunctuated ASR text.

A plausible implication is that Chuan-Pipeline is designed not merely to maximize transcript correctness, but to preserve a training signal that is simultaneously useful for recognition, synthesis, and speaker-conditioned modeling. That orientation distinguishes it from pipelines optimized only for raw ASR throughput.

4. Evaluation benchmarks

WenetSpeech-Chuan-Eval consists of manually verified ASR and TTS benchmarks designed to expose both ordinary and difficult dialectal conditions. The ASR benchmark, WSC-Eval-ASR, totals 9.7 hours and is divided into an Easy subset of 8.55 hours and a Hard subset of 1.15 hours (Dai et al., 22 Sep 2025).

Benchmark subset	Main domain(s)	Duration (h)
WSC-Eval-ASR Easy	audiobooks, reading	8.55
WSC-Eval-ASR Hard	short videos, entertainment, drama	1.15
WSC-Eval-ASR Total	/	9.7

The Easy portion emphasizes relatively clean, read speech, while the Hard portion emphasizes spontaneous or noisy media speech. All ASR evaluation samples retain age, gender, and emotion labels, enabling fine-grained analysis beyond a single corpus-level CER.

The TTS benchmark, WSC-Eval-TTS, is divided into easy and hard conditions as well. WSC-Eval-TTS-easy contains sentences with Sichuanese dialectal words across diverse domains. WSC-Eval-TTS-hard includes long sentences and LLM-generated sentences such as tongue twisters, folk sayings, and emotional speech. For prompt conditioning and speaker-similarity evaluation, the benchmark uses 10 speakers—5 male and 5 female—from MagicData and internal recordings, with 200 sentences per speaker (Dai et al., 22 Sep 2025).

ASR is evaluated primarily with CER, while TTS is evaluated with both objective and subjective metrics. Objective TTS metrics are CER—computed by recognizing synthesized audio with FireRedASR—and SIM, a speaker-similarity score computed with WavLM-Large. Subjective metrics are IMOS for intelligibility, SMOS for speaker similarity, and AMOS for accent naturalness. AMOS is judged by 10 native Sichuanese raters and 10 non-expert listeners over 30 samples (Dai et al., 22 Sep 2025).

5. Benchmark models and reported performance

The paper evaluates both ASR and TTS systems trained or fine-tuned on WenetSpeech-Chuan. On the ASR side, the key systems include Paraformer (220M), FireRedASR-AED (1.1B), SenseVoice-small (234M), Whisper-small (244M), Kimi-Audio (7B), and Qwen2.5-omni (3B), together with WenetSpeech-Chuan-finetuned variants (Dai et al., 22 Sep 2025).

The clearest ASR result is that Paraformer-WSC-Finetune reduces average CER from 16.52 to 14.58, described as an improvement of about 11.7% over baseline Paraformer. A further variant, Paraformer-WSC-Finetune + internal data, reaches 12.87 CER on WSC-Eval-ASR Total, 14.59 on MagicData Conversation, 6.28 on MagicData Daily-Use, and 13.38 average CER, which the paper characterizes as state-of-the-art among open-source models and comparable to commercial systems (Dai et al., 22 Sep 2025).

Fine-tuning also benefits omni-modal models. Qwen2.5-omni-WSC-Finetune improves average CER from 17.69 to 15.74, roughly 11% better than the base omni model. By contrast, generic multilingual or general speech models without Sichuanese specialization, such as Whisper-small and SenseVoice-small, are reported to perform much worse on the dialectal benchmarks (Dai et al., 22 Sep 2025).

On the TTS side, CosyVoice2-WSC—with its LLM component trained on the full WenetSpeech-Chuan corpus—achieves strong dialectal synthesis. On WSC-Eval-TTS-easy it reports CER 4.28, SIM 72.78, IMOS 4.13, SMOS 3.94, and AMOS 4.05. A further supervised fine-tuned model, CosyVoice2-WSC-SFT, reaches CER 4.08, SIM 78.84, SMOS 4.16, and AMOS 4.20 on the easy set, and CER 7.22, SIM 67.96, IMOS 4.01, SMOS 3.03, and AMOS 3.98 on the hard set (Dai et al., 22 Sep 2025).

A notable pattern is that WenetSpeech-Chuan-trained TTS systems are not merely intelligible; they are competitive on accent naturalness. On hard TTS evaluation, CosyVoice2-WSC and CosyVoice2-WSC-SFT achieve AMOS 3.92 and 3.98, respectively, slightly above the reported 3.88 for commercial Qwen-TTS. This suggests that the corpus is especially valuable for dialectal accent control rather than only for neutral-text reconstruction.

6. Research significance, interpretation, and release

WenetSpeech-Chuan addresses a structural bottleneck in dialectal speech processing: the absence of large, openly accessible corpora that capture real dialect use rather than small read-speech datasets or accented-Mandarin approximations. In that sense, its most important contribution is not only the 10,013-hour scale, but the combination of authentic Sichuanese, rich paralinguistic annotation, and manually verified ASR/TTS benchmarks (Dai et al., 22 Sep 2025).

A common simplification is to treat Sichuanese as merely a pronunciation variant of Standard Mandarin. The corpus design directly contradicts that simplification by targeting dialectal differences in tone, lexicon, and grammar, and by evaluating on content such as folk sayings, dialectal vocabulary, and emotionally expressive speech. This also clarifies why generic Mandarin or multilingual speech models degrade on the benchmark and why dialect-specific fine-tuning yields large gains.

The release is also framed as an AI equity intervention. The paper argues that underrepresentation of dialect speakers in mainstream ASR and TTS systems produces a systematic performance gap, and that an open 10,000-hour Sichuanese corpus helps reduce that bias. An important empirical point is that the reported dialect gains are said not to come at the cost of Mandarin performance, although the detailed Mandarin-side results are deferred to the project page (Dai et al., 22 Sep 2025).

The corpus, benchmarks, models, and recipes are released through the project repository at https://github.com/ASLP-lab/WenetSpeech-Chuan (Dai et al., 22 Sep 2025). A plausible broader implication is that WenetSpeech-Chuan establishes a template for future dialect corpora: large-scale web mining, diarization and speaker clustering, LLM-assisted transcript correction, multimodal punctuation, rich speaker metadata, and evaluation sets deliberately split by difficulty. As such, it is likely to function not only as a Sichuanese benchmark, but also as a methodological reference point for dialectal speech-resource construction in Chinese and beyond.