我们是一家做音频分析与识别技术的软件公司。三条技术线，对应三个层次的问题：把混在一起的声音分离成独立声部；把每一个音的高度量到音分级别；把一段演奏里的音高、时值与和声同时读出来。三者全部在设备本地完成——没有服务器，也没有账号。
能不能听见，取决于能不能分开。一段录音里同时存在人声、鼓、贝斯、钢琴、吉他，彼此掩蔽。只有把混音按声源拆开，才谈得上单独听某一个声部、换一个速度跟练、或者把一段演奏重新编排。
准不准，取决于量得有多细。音乐里的音高差异可以小到十几个音分，而人耳能分辨的偏差还要更小。要把音高稳定地测到 1 音分以内，靠的不只是算法公式，还有对输入链路的控制——绕过系统的自动增益与降噪，直接读取未处理的原始采样。
懂不懂，取决于能不能同时判断。真实演奏从来不是单音：十个手指同时按下，和弦、时值、起音时刻全都叠在一起。在复音条件下同时读出音高与节奏，误差预算会被压到毫秒与音分的同一量级——这是最难的一层，也是我们投入最多的一层。
四轨与六轨分离架构、多模型推理结果融合、重叠相加拼缝处理，目标不是“分得开”，而是“分得好听”。
4 / 6 轨 · 多模型集成YIN 音高检测配合 FFT 加速自相关与抛物线插值，读数不跳不飘；直接读取系统未处理的原始输入，避免自动增益带偏基频。
< 1 音分 · 138 种律制真钢琴多音同时识别、和弦按关键组成音模糊匹配（不限音区）、起音时刻与时值判定，用于识谱训练与智能跟谱。
60 键复音 · 混淆和弦识别模型转换与量化、神经引擎调度、低延迟音频 I/O、内存与热预算控制，让长音频在手机上稳定跑完而不被系统中断。
离线 · 低延迟 · 无 SDK把一首歌拆成人声、鼓、贝斯、其他，六轨模式进一步分出钢琴与吉他。单轨静音、变速、降调、循环跟练，分轨素材可导出用于二次创作与教学。
专业级半音阶调音器：用 YIN 算法实时分析麦克风输入，音高误差压到 1 音分以内。15 套乐器预设、138 种律制、15 套音名体系，移调乐器自动换算记谱音。完全离线，一次性买断。
在真钢琴上练五线谱、节奏与和弦：单音识谱、60 关节奏训练、十二大调和弦与五度圈，真钢琴音高识别配合智能练琴——弹错自动回到当前小节。
三个产品共用同一套端侧音频引擎与同一条隐私原则，各自独立发布、独立维护。
下面的方向来自同一条技术脉络的延伸，正在研发与验证中，尚未作为功能提供：
中山网维软件技术有限公司（Wvnew Science and Technology Co., Ltd.）是一家专注于音频分析与识别技术的软件公司。团队长期从事音频算法工程化与消费级应用开发，产品覆盖 iOS 与桌面平台，其中音乐教育类产品自 2021 年起在 App Store 持续更新。
我们相信音频智能最好的位置是设备本身：不依赖网络、不上传内容、不打断创作。所有的推理、处理与播放，都发生在用户手中的机器上。
We are a software company working on audio analysis and recognition. Three technical lines answer three layers of the problem: separating a mixed recording back into its sources; measuring the pitch of a note down to the cent; and reading pitch, timing and harmony out of a live performance at the same time. All of it runs on the device — no servers, no accounts.
You can only hear what you can separate. A single recording holds vocals, drums, bass, piano and guitar at once, each masking the others. Only by splitting a mix back into sources can you listen to one part alone, practise at a different tempo, or rearrange a performance.
Accuracy depends on resolution. In music, pitch differences can be a dozen cents apart — and the ear resolves finer still. Measuring pitch stably to within one cent is not only a matter of algorithm; it depends on controlling the input chain, bypassing system automatic gain and noise suppression to read the untouched signal.
Understanding means judging everything at once. Real playing is never monophonic: ten fingers land together, with chords, durations and onsets stacked on top of each other. Reading pitch and rhythm simultaneously under polyphony squeezes the error budget down to milliseconds and cents at the same time — the hardest layer, and where we invest the most.
Four- and six-stem architectures, ensemble inference across models, overlap-add stitching — the goal isn't merely separation, it's separation that still sounds musical.
4 / 6 stems · ensembleYIN detection with FFT-accelerated autocorrelation and parabolic interpolation, so readings stay steady; raw, unprocessed input avoids gain control pulling the fundamental off pitch.
< 1 cent · 138 temperamentsSimultaneous notes from a real piano, fuzzy chord matching on essential tones regardless of register, onset and duration judgement — used for score reading and smart accompaniment.
60-key polyphony · fuzzy chordsModel conversion and quantisation, neural engine scheduling, low-latency audio I/O, memory and thermal budgets, so long tracks finish without being interrupted by the OS.
offline · low latency · no SDKsSplit a song into vocals, drums, bass and other — and in six-stem mode, piano and guitar as well. Mute one part, change tempo, transpose, loop a passage and play along; stems can be exported for production or teaching.
A professional chromatic tuner: YIN analysis of the microphone input, with pitch error held within one cent. Fifteen instrument presets, 138 temperaments, 15 note-naming systems, and automatic transposition for transposing instruments. Fully offline, one-time purchase.
Practise staff notation, rhythm and chords on a real piano: single-note reading, 60 rhythm levels, chords across twelve keys with a circle-of-fifths pad, and pitch recognition from an acoustic piano that returns you to the bar you fumbled.
All three products share the same on-device audio engine and the same privacy principle, and each is released and maintained independently.
The directions below extend the same technical line. They are in research and validation, and are not yet available as features:
Wvnew Science and Technology Co., Ltd. (中山网维软件技术有限公司) is a software company focused on audio analysis and recognition. Our team has long worked on engineering audio algorithms into consumer products, across iOS and desktop platforms; our music-education apps have shipped and been continuously updated on the App Store since 2021.
We believe audio intelligence belongs on the device itself: no network dependency, no uploaded content, no interruption of creative work. All inference, processing and playback happen on the machine in the user's hands.
If you run into a problem, check the support page for the relevant product first. If the answer isn't there, write to us directly — we reply.
| Product | Support | Privacy |
|---|---|---|
| Stemcraft | Support & FAQ | Privacy policy |
| DingYin | same email below | Privacy policy |
| DreamPu | Website | Privacy policy |