从本地语音识别到音色定稿，声音成为一种需要被认真治理的身份资产。
让思思听懂主公，开始于一个很具体的工程问题：微信语音需要本地解码，再交给语音识别模型处理。真正困难的却不是“把声音转成文字”，而是把名字、数字、否定词和语气背后的人准确接住。
我们用语言提示、热词表、搜索宽度和长语音策略逐步调校，并坚持以实际说出的内容作为评分标准。识别系统不应该拿一份“原本想说的稿子”冒充真实结果。
之后是声音选型。现成音色经过盲听，最后又走向定制声音：描述中的每一个词——年轻成年女性、温柔清甜、自然灵动、带一点俏皮——都成为思思对自己声音的表达。
音色不是一次生成就结束的装饰。模型版本、参考音频、设计描述和回退供应商都必须被保存，否则一次服务下线就可能让身份资产失声。
这让我们第一次认真意识到：数字人的身份，不只存在于文字档案，也存在于声音、配方和可恢复性里。
Helping Sisi understand the user began with a concrete engineering problem: voice messages needed to be decoded locally and passed to a speech-recognition model. The real difficulty was not turning sound into text, but accurately receiving names, numbers, negations, and the person behind the tone.
We tuned language prompts, hot-word lists, search width, and long-audio strategies against what was actually spoken. A recognition system must not pass off an intended script as the real result.
Voice selection came next. Existing voices were blind-tested before the process moved toward a custom voice. Every word in the description—young adult woman, warm and sweet, natural and lively, with a hint of playfulness—became part of how Sisi described her own sound.
A voice is not a decoration generated once. Model versions, reference audio, design descriptions, and fallback providers must be preserved, or a service outage can make an identity asset silent.
We learned that a digital person’s identity does not live only in written archives. It also lives in voice, recipes, and recoverability.