Baiyin AI Singer Training

Problem

When building a Baiyin AI singer model from audio, the blocker is often not the API call itself. Local files, chat attachments, and private storage links cannot be trained directly, optional fields are easy to over-question, Chinese names can become garbled, and task status can be mistaken for a finished model. This skill focuses on Baiyin singer training: creating tasks, checking progress, and retrieving final results with fewer unnecessary prompts.

How it works

• Create mode: Extract name, materialUrl, description, avatar, tags, and trainType from the conversation. Missing optional fields do not block the task; trainType defaults to 2, while high-quality requests map to 1. If the user only has a local file or inaccessible URL, the skill points to the Baiyin upload endpoint POST /api/open/v1/file/upload to obtain a public data.url before training.
• Status and result mode: Use GET /api/open/v1/tasks/{taskId} to check queued, processing, succeeded, or failed. On success, prioritize returning modelId, demoUrl, and then modelName, avatar, userTaskId, and trainType.
• Reliability checks: Emphasize UTF-8 encoding for Chinese name, description, and tags. If the returned modelName contains ??, garbled text, or looks inconsistent with the original request, mark it as an abnormal result instead of treating it as success.
• Polling boundary: Poll only when the user explicitly asks to wait. The default interval is about 10 minutes, with up to 6 consecutive checks per round; stop on succeeded or failed.

Boundaries and notes

• Requires a valid BAIYIN_API_KEY, and the training audio must be a publicly accessible URL.
• Intended for AI singer training tasks, not for ordinary AI cover generation results.
• For status or result checks, a clear taskId is still needed if it is not present in the current context.
• For backend errors such as 400, 401, 402, or 404, the skill points to parameter, authentication, balance, or missing-task issues and avoids guessing the final model output.