Chuan-Pipeline: Sichuanese Speech Corpus
- Chuan-Pipeline is a comprehensive framework that transforms raw multi-domain audio into a richly annotated Sichuanese speech corpus, addressing data scarcity and quality issues.
- It employs a sequential architecture with stages such as segmentation, speaker diarization, LLM-based transcription correction, and multi-modal punctuation prediction to ensure reliability.
- The framework supports robust ASR and TTS benchmarks by partitioning data into strong- and weak-label sets and incorporating manual verification for benchmark-grade evaluation.
Chuan-Pipeline is a comprehensive, systematic data processing framework developed to create large-scale, high-quality speech corpora for the Sichuanese (Chuan) dialect. Introduced with "WenetSpeech-Chuan: A Large-Scale Sichuanese Corpus with Rich Annotation for Dialectal Speech Processing" (Dai et al., 22 Sep 2025), its principal function is to transform massive amounts of raw, unlabeled, multi-domain audio data into a richly annotated corpus, thereby addressing the scarcity and quality issues that hinder dialectal speech resources. In the reported instantiation, the framework underlies the construction of WenetSpeech-Chuan, a 10,013-hour corpus partitioned into strong-label and weak-label subsets, and supports both ASR and TTS benchmarks with manually verified transcriptions (Dai et al., 22 Sep 2025).
1. Provenance and problem setting
Chuan-Pipeline was proposed to address a specific bottleneck in dialectal speech technology: the scarcity of large-scale, open-source data for dialects, especially for the widely spoken Sichuanese dialects of Chinese. The framework is presented as a complete data processing framework for dialectal speech, rather than as a narrow transcription component or a benchmark-only curation tool. Its scope spans raw-data acquisition, segmentation, speaker processing, paralinguistic annotation, automatic quality assessment, transcription generation and correction, punctuation restoration, manual verification for evaluation sets, and final corpus construction (Dai et al., 22 Sep 2025).
The reported motivation is twofold. First, dialectal speech exhibits high pronunciation and phonetic variation, weak or noisy supervision, and a lack of punctuation in spontaneous audio. Second, dialectal corpora often lack speaker and paralinguistic metadata that would support robust recognition and synthesis. Chuan-Pipeline addresses these issues by combining automated filtering and annotation with human verification at selected stages, particularly for benchmark construction (Dai et al., 22 Sep 2025).
A plausible implication is that the framework is intended not merely to enlarge a dataset, but to formalize a repeatable dialect-resource construction workflow. That interpretation is consistent with the paper’s characterization of the pipeline as a complete and systematic framework and with the release of code, recipes, models, and benchmarks on the project page (Dai et al., 22 Sep 2025).
2. Sequential architecture
The architecture is described as a sequence of six main stages. The sequence is linear at the level of corpus construction, but several stages incorporate iterative refinement or human feedback.
| Stage | Main operation | Reported outcome |
|---|---|---|
| 1 | Pre-Processing and Labeling | Segmented, speaker-processed, paralinguistically annotated clips |
| 2 | Quality Assessment | Low-quality segments discarded via duration, SNR, and WVMOS |
| 3 | LLM-GER Transcription | Multiple ASR hypotheses merged and corrected |
| 4 | Punctuation Prediction | Audio-text fusion adds punctuation aligned with speech rhythm |
| 5 | Manual Verification | Gold-standard ASR/TTS evaluation data |
| 6 | Corpus Construction & Partitioning | Confidence-based strong and weak labels |
The first stage begins with data acquisition by crawling audio from online domains such as short videos, live-streams, and entertainment. This is followed by an initial dialect verification, described as a manual check to ensure the presence of Sichuanese dialect. Long audios are then segmented with voice activity detection into clips of 5–25 seconds, with silence and background noises removed. Single-speaker selection is performed with the pyannote toolkit for speaker diarization, and speaker embedding and clustering are conducted with CAM++, which extracts speaker vectors and supports consistent speaker IDs (Dai et al., 22 Sep 2025).
The same stage also performs paralinguistic annotation. Speaker gender is inferred with voice-gender-classifier, age with Vox-Profile, and emotion through majority vote over Emotion2vec and SenseVoice, with accuracies up to 98.7%. The reported age categories are children, teenager, young, middle-aged, and old. The emotion categories are happy, angry, sad, neutral, fearful, surprised, and disgusted (Dai et al., 22 Sep 2025).
Subsequent stages focus on quality, labeling, and corpus usability. Automatic quality assessment scores segments using duration, SNR, and word-level virtual MOS. LLM-GER transcription then fuses outputs from multiple ASR engines. A multi-modal punctuation model integrates text with pause structure derived from forced alignment. Manual verification is reserved for the evaluation subsets, while the full corpus is partitioned by confidence into strong-label and weak-label segments (Dai et al., 22 Sep 2025).
3. Core processing mechanisms
A central technical feature of Chuan-Pipeline is its multi-system transcription strategy, termed LLM-GER, or LLM - Generative Error Correction-based ROVER. Three independent ASR engines—FireRed-ASR, SenseVoice-Small, and TeleASR—first generate separate hypotheses for each utterance. Qwen3 then merges and corrects these outputs with a custom prompt designed to preserve both semantic integrity and token correspondence. The description characterizes this as conceptually reflecting ROVER while replacing token-level voting with generative LLM correction, and reports improved transcription accuracy by approximately 15% over individual ASR systems (Dai et al., 22 Sep 2025).
The paper’s detailed description summarizes the transcription stage with the following abstractions: and
where the confidence function is described conceptually in terms of token-wise agreement, edit distance, or related agreement structure. The final transcription is the merged hypothesis when confidence is high; otherwise the segment may be excluded or flagged for further review. This suggests that confidence is not an auxiliary score added after transcription, but an operational variable that affects retention and downstream labeling (Dai et al., 22 Sep 2025).
The punctuation module is explicitly multi-modal. The motivation given is that text-only punctuation prediction fails to capture authentic speech rhythm. Kaldi is used for forced alignment to obtain word timestamps and inter-word pause durations. Pause analysis distinguishes short and long pauses using thresholds such as 0.25s for short and 0.5s for long, with thresholds iteratively optimized. A BiLSTM-based punctuation predictor then maps pause candidates to punctuation marks: a short pause to a comma, and a long pause to a period, question mark, or exclamation mark. Human-in-the-loop feedback is used to refine thresholding (Dai et al., 22 Sep 2025).
Speaker processing is another key algorithmic layer. CAM++ extracts speaker-level features, represented in the description as , and clustering assigns utterances to distinct speakers. In combination with pyannote-based diarization, this supports single-speaker segment selection and stable speaker attribution across clips. Because the framework is intended for both ASR and TTS, this step is structurally important: a plausible implication is that it reduces cross-speaker contamination not only in recognition training but also in synthesis training (Dai et al., 22 Sep 2025).
4. Quality control and corpus partitioning
Quality control in Chuan-Pipeline is based on automatic assessment followed by selective pruning. Each segment is evaluated for duration and SNR, and is also assigned a word-level virtual MOS. The detailed description states that WVMOS is a computed proxy for perceptual audio quality using duration and SNR features, and summarizes it as
with described as a regression model trained to estimate perceptual MOS given acoustic metrics. Segments with low WVMOS are discarded, and most retained audio is reported to fall in the range 3.0–3.5 on the MOS scale (Dai et al., 22 Sep 2025).
This stage is directly tied to the problem of weak or noisy supervision. The framework’s stated response is to use WVMOS, confidence scores, and selective pruning to manage uneven recording and transcription quality. In other words, quality control is distributed across both the acoustic and textual channels: acoustic quality is screened by WVMOS and SNR, while transcription reliability is screened by cross-system agreement and LLM-GER confidence (Dai et al., 22 Sep 2025).
The final partitioning is confidence-based. Each audio segment receives a confidence score based on transcription reliability, after which the corpus is divided into a strong-label subset and a weak-label subset. The reported thresholds and durations are:
| Partition | Confidence rule | Hours |
|---|---|---|
| Strong Label | Confidence | 3,714 hrs |
| Weak Label | $0.60 <$ Confidence | 6,299 hrs |
| Total | — | 10,013 hrs |
This partition is significant because it embeds uncertainty into the dataset structure rather than collapsing all automatically transcribed speech into a single label regime. A plausible implication is that the corpus can support training strategies that treat strong and weak supervision differently, although that specific training strategy is not detailed in the provided description (Dai et al., 22 Sep 2025).
5. Benchmark construction and empirical role
Chuan-Pipeline is not limited to large-scale corpus generation; it also provides the preparation logic for evaluation benchmarks. For WSC-Eval-ASR and WSC-Eval-TTS, all transcriptions are manually checked and corrected by professional annotators, and speaker attributes including age, gender, and emotion are also annotated. This step is explicitly framed as necessary to ensure gold-standard benchmark quality (Dai et al., 22 Sep 2025).
The ASR evaluation set, WSC-Eval-ASR, is reported as 9.7 hours, split into Easy and Hard subsets, multi-domain, manually verified, and equipped with rich speaker annotation. The TTS evaluation set, WSC-Eval-TTS, is described as having easy and hard splits, ten speakers with 5 male and 5 female, and coverage of both regular and challenging sentences and styles (Dai et al., 22 Sep 2025).
The paper further states that models trained on WenetSpeech-Chuan achieve state-of-the-art performance among open-source systems and demonstrate results comparable to commercial services, and that Chuan-Pipeline supports state-of-the-art CER for ASR and MOS for TTS in open-source benchmarks (Dai et al., 22 Sep 2025). Within the confines of the provided description, the significance is methodological as much as empirical: the pipeline is presented as the enabling mechanism by which benchmark-grade evaluation data and large-scale training data coexist within a unified dialect-resource construction framework.
6. Significance, scope, and terminological clarification
Within dialectal speech processing, Chuan-Pipeline is significant because it addresses four challenges identified in the paper: high pronunciation and phonetic variation, weak or noisy supervision, lack of punctuation, and speaker or paralinguistic diversity. It does so by combining multiple ASR hypotheses with LLM-based correction, audio-informed punctuation prediction, automated multi-label speaker annotation, and confidence-aware data selection (Dai et al., 22 Sep 2025). The corpus, benchmarks, models, and receipts are publicly available on the project page, and the open-source release includes Chuan-Pipeline, annotated corpus assets, benchmark recipes, trained models, custom LLM prompts for GER, and annotation scripts (Dai et al., 22 Sep 2025).
A common misconception is to treat the name merely as a generic “pipeline” label. In the cited speech-processing usage, Chuan-Pipeline denotes a dialect-specific corpus-construction framework for Sichuanese speech, not a distributed training schedule, a database execution scheme, or an accelerator mapping strategy. It is therefore unrelated in subject matter to pipeline-parallel training frameworks such as "Pipeline Parallelism with Controllable Memory" (Qi et al., 2024), "PipeTransformer: Automated Elastic Pipelining for Distributed Training of Transformers" (He et al., 2021), or "Pipeline MoE: A Flexible MoE Implementation with Pipeline Parallelism" (Chen et al., 2023), which address model-parallel optimization rather than speech-resource curation.
Another possible source of ambiguity is nominal rather than conceptual. In the provided literature, a distinct 2026 study on vertical two-phase flow also uses the label “Chuan-Pipeline” for an unsupervised topological regime-discovery framework based on Euler Characteristic Surfaces and Multiple Kernel Learning (Koenig et al., 7 Apr 2026). That usage concerns churn-flow characterization in small-diameter vertical pipes and is unrelated to Sichuanese ASR or TTS. This suggests that the term is not globally unique across arXiv literature and should be interpreted in context.
In its primary, speech-technology sense, however, Chuan-Pipeline denotes a domain-specific, multi-stage data processing pipeline for dialectal speech corpora. Its reported contribution lies in orchestrating data scraping, segmentation, speaker and paralinguistic annotation, LLM-based transcription correction, multi-modal punctuation restoration, and rigorous quality control to make large-scale Sichuanese ASR and TTS research technically tractable and reproducible (Dai et al., 22 Sep 2025).
