線上音樂轉換器Online Music Converter
在瀏覽器中完成：去除人聲、樂器解析、播放、音樂檔與樂譜輸出。音樂檔不會上傳到任何伺服器。Everything runs in your browser: vocal removal, instrument analysis, playback, audio and sheet-music export. Your audio is never uploaded.
1輸入音樂檔Load a music file
（預設關閉）上傳同一首歌的含人聲版與伴奏版，兩版本對齊相減，取得精確人聲並更準確解析樂器。(off by default) Upload the vocal and instrumental versions of the same song; they are aligned and subtracted for exact vocals and more accurate instrument analysis.
含人聲版（原曲）Vocal version (full mix)
伴奏版（Instrumental）Instrumental version
點擊選擇或將音樂檔拖曳到此處Click to choose or drop an audio file here
MP3 / WAV / M4A / OGG / FLAC
2去除人聲Remove vocals
自動：先以訊號處理快速分離（依每首歌的立體聲特性調整強度與人聲純化），再於背景以 AI 模型（UVR MDX-Net）重新分離，完成後自動替換人聲與伴奏，人聲更乾淨、無雜音。首次使用需下載約 67 MB 模型（之後瀏覽器會快取）；處理時間視電腦效能而定：支援 WebGPU 的瀏覽器（Chrome、Edge）最快，否則以多核心 CPU 平行運算，一首 4 分鐘的歌約需數分鐘至十餘分鐘（進度列會顯示剩餘時間）。等待期間可先播放原曲。單聲道檔案亦可用 AI 分離。Auto: a fast signal-processing split first (strength and purification tuned to the song's stereo mix), then an AI model (UVR MDX-Net) re-separates in the background and replaces the vocals and instrumental when done: cleaner vocals without artefacts. The first run downloads a ~67 MB model (cached afterwards). Speed depends on the computer: fastest with WebGPU (Chrome, Edge), otherwise it runs on several CPU cores in parallel; a 4-minute song takes from a few minutes to over ten (the progress bar shows the remaining time). You can play the original meanwhile. Works on mono files too.
（追蹤歌聲音高，只保留人聲泛音，濾除置中的伴奏樂器與鼓聲）(tracks the singing pitch and keeps only the voice's harmonics, removing centred instruments and drums)
原理：以短時傅立葉轉換比對左右聲道，削減置中且相似的成分（多數歌曲的主唱位置）；再以「人聲純化」追蹤歌聲音高，把同樣置中的鋼琴、貝斯、鼓等非人聲成分退回伴奏。單聲道檔案無法以此方式分離。Method: STFT left/right similarity masking removes centre-panned content in the vocal band (where lead vocals usually sit); "vocal purification" then tracks the singing pitch and returns other centred instruments (piano, bass, drums) to the instrumental. Mono files cannot be separated this way.
分析模式：人聲 = 含人聲版 − 伴奏版（對齊後精確相減），不需調整上方參數。Analysis mode: vocals = vocal version − instrumental version (aligned, exact subtraction). The settings above are not used.
兩版本若母帶、等化或壓縮不同，相減後會殘留音樂。此處依頻帶與時間自動匹配音量，再扣除殘留並在人聲間隔自動靜音；數值越高越乾淨，但過高可能削弱細微人聲（如氣音、尾音）。If the two versions differ in mastering, EQ or compression, subtraction leaves some music. This matches levels per band over time, suppresses the leak and mutes the gaps between phrases. Higher is cleaner, but too high can thin out soft vocal details (breaths, tails).
3播放Player
00:00.00 / 00:00.00
A – B: –
（開啟後，第 6 欄樂譜上會有符號跟著目前播放的音符跳動）(a symbol hops along the score in section 6 with the note being played)
時間軸可點擊或拖曳定位（精確到 0.01 秒），滑鼠滾輪可縮放。格放 / 格退會移動一格並試聽該瞬間。Click or drag the timeline to seek (0.01 s precision); scroll to zoom. Frame fwd / back moves one step and auditions that instant.
4樂器解析Instrument analysis
長度Duration–
速度Tempo–
調性Key–
聲道Channels–
兩版本對齊位移Alignment offset–
伴奏吻合度Instrumental match–
人聲出現比例Vocal presence–
人聲音域Vocal range–
推估聲部Voice type (est.)–
和聲 / 疊唱Backing vocals–
✓ 偵測到Detected
~ 可能 / 相關Possible / related
· 未偵測Not detected
以頻譜特徵（頻帶能量、起音、和聲性、立體聲位置）估計，屬啟發式分析，非經訓練的辨識模型；點擊「獨奏」可試聽該樂器所在頻帶。Estimated from spectral features (band energy, onsets, harmonicity, stereo position). This is heuristic analysis, not a trained classifier. Press "Solo" to audition the band where an instrument sits.
6輸出樂譜Sheet music export
（不懂設定？開啟後只要選擇樂器與譜式，即自動以人聲主旋律產生可彈奏的樂譜）(Not sure about the settings? Turn this on, pick an instrument and notation, and a playable score is made from the vocal melody.)
MusicXML 可用 MuseScore、Finale、Sibelius 開啟編輯；自動採譜為近似結果，建議再人工校正。MusicXML opens in MuseScore, Finale or Sibelius. Automatic transcription is approximate; proofread before use.
7轉譯歌詞Lyrics transcription
使用開源語音辨識模型 Whisper，在您的瀏覽器內執行（首次需下載模型，之後會快取）；音樂不會上傳。歌聲辨識難度高於說話，建議先用分析模式取得乾淨人聲，結果仍需校對。歌詞一律以「歌詞語言」辨識；第二、三語言作為備援（某段以主要語言辨識不出內容時改用）。翻譯優先使用 Chrome 內建的離線翻譯（Chrome 138 以上，免下載）；否則改用開源翻譯模型 NLLB（首次約需下載 600 MB 以上）。Uses the open-source Whisper model running in your browser (downloaded once, then cached); your audio is not uploaded. Sung lyrics are harder than speech; clean vocals from analysis mode help, and results still need proofreading. Lyrics are always recognised in the "lyrics language"; the 2nd/3rd languages are fallbacks for segments where nothing is recognised. Translation uses Chrome's built-in on-device translator (Chrome 138+, no download); otherwise the open NLLB model (first use downloads 600 MB or more).
（以網路公開歌詞資料庫校正文字並對齊歌聲）(correct the text with public online lyrics and align it to the singing)
或直接貼上歌詞（純文字，或含時間碼的 LRC 格式）Or paste lyrics (plain text, or LRC with timestamps)
歌詞來源為公開資料庫 LRCLIB（lrclib.net），由您的瀏覽器直接查詢；含時間碼的歌詞會自動計算時間偏移並對準每句開口處，純文字歌詞則與辨識結果逐句比對。歌詞版權屬原作者，僅供個人學習使用。Lyrics come from the public LRCLIB database (lrclib.net), queried directly by your browser. Time-stamped lyrics get an automatic offset and are snapped to each sung phrase; plain lyrics are matched line by line against the recognised text. Lyrics are copyrighted by their owners; personal study use only.
© 2026 AEML, NTUST
國立臺灣科技大學 營建工程系National Taiwan University of Science and Technology · v2.5
No comments:
Post a Comment