A compact, studio-recorded dataset for rapid evaluation and adaptation of TTS, speech-to-speech, ASR, and synthetic-speech detection systems.
Two focused collections give your model something harder to source than clean read speech: controlled emotion and natural conversational rhythm.
Anger, happy, sad, fear, neutral, disgust, and coquetry across three voices for expressive TTS and speech-to-speech evaluation.
Long and short turns with pacing, hesitation, emphasis, and everyday delivery for conversational agents and voice UX.
The first release focuses on two consistent studio voices rather than mixing every asset in the library. Each production candidate is paired with text and a stable ID.
于是狐狸踩着山羊的背跳出了井口，他马上就准备溜走。
麻烦您过来一下獾子先生，您帮忙把老鼠弹弓的这头往下拉住。
日前，见到治污工程已基本完工，但部分混凝土尚未完全凝固。
虎、美洲虎等森林大猫靠伏击捕猎，而大草原上不适合搞伏击。
Original WAV masters plus task-specific resampling on request.
Simplified Chinese transcripts with source and delivery IDs.
Speaker, duration, technical fields, checksums, and QC results.
Data card, schema, evaluation terms, and commercial license scope.
The source collection is owner-confirmed as commercially licensable. Every external delivery is paired with a written project license and an indexed evidence package.
Traceable sourceStable source paths, clip IDs, and file hashes support provenance review.
Scoped permissionTraining, fine-tuning, evaluation, term, territory, and deployment are defined in writing.
No implied voice clone rightsIdentifiable synthetic voice use is licensed separately when required.
No resale of raw dataCustomer use is limited to the agreed model or project scope.
Owner-confirmed commercially licensable
Evidence documents indexed before external delivery
Fine-tune pronunciation, pacing, and naturalness without commissioning a new studio collection.
Evaluate Mandarin output quality for conversational agents and expressive voice systems.
Test recognition performance on clean studio speech, long-form lines, and varied sentence structures.
Use verified human recordings as a clean reference set for detection and robustness experiments.
Start with a controlled evaluation. Expand only when the data improves the model.
Review a 100-clip sample with transcripts, metadata, hashes, and QC fields.
Map the files to your pipeline and compare against a defined Mandarin baseline.
License the 18.825-hour candidate pool for one scoped model or project.
Add speakers, styles, annotations, or custom studio recording after measurable improvement.
30–90 minutes, complete metadata, 30-day internal evaluation.
One scoped model or project, non-exclusive, 6–12 month pilot term.
Formatting, baseline comparison, small-scale fine-tuning, and results report.
Indicative pilot ranges. Final terms depend on model scope, deployment rights, term, and required annotations.
Includes WAV files, CSV and JSONL manifests, transcripts, SHA-256 hashes, acoustic QC fields, a readme, and evaluation-only terms.
Not by default. Identifiable synthetic voice generation requires an explicit license scope beyond general model evaluation or adaptation.
Yes. We retain 48 kHz masters and can provide task-specific derivatives while preserving the original source files and provenance.
Yes. Additional speakers, styles, emotional speech, annotations, and custom studio recording can be scoped after the pilot.
No. The short previews above are public, while the complete evaluation pack remains access-controlled so the recipient and intended use are documented before delivery.
Request the evaluation pack, tell us the target model and use case, and receive a scoped pilot proposal.