music-lens 0.1.2 → 0.1.4

package/README.md

@@ -23,10 +23,14 @@ music-lens analyze song.mp3
 ```powershell
 $node = "C:\Users\86173\Documents\deepseek-harness\default-workspace\snowluma\SnowLuma-v2.2.1\node.exe"
 & $node .\bin\music-lens.mjs doctor                       # 体检：python / 依赖 / 模型
+& $node .\bin\music-lens.mjs spectrogram "song.mp3"       # 10 秒出图，只要 ffmpeg
 & $node .\bin\music-lens.mjs separate "song.mp3"          # 分轨
 & $node .\bin\music-lens.mjs analyze "song.mp3" --model medium
 ```
 
+`spectrogram` 是**零依赖**的快速预览：频谱图 + 波形图，不装 Python、不下模型，
+只要机器上有 ffmpeg（`MUSIC_LENS_FFMPEG` 可指定路径）。
+
 `setup` 会自己装 python 依赖并从 hf-mirror 下模型，**默认只下 small（461 MB）**：
 
 ```powershell

package/bin/music-lens.mjs

@@ -77,6 +77,22 @@ function findPython() {
   return null
 }
 
+// ffmpeg is optional: only the zero-dependency `spectrogram` command needs it.
+function findFfmpeg() {
+  const cands = []
+  if (process.env.MUSIC_LENS_FFMPEG) cands.push(process.env.MUSIC_LENS_FFMPEG)
+  cands.push('ffmpeg')
+  if (process.platform === 'win32') {
+    cands.push(path.join(os.homedir(), '.dsh', 'bili-whale', 'bin', 'ffmpeg.exe'))
+    cands.push('C:\\ffmpeg\\bin\\ffmpeg.exe')
+  }
+  for (const c of cands) {
+    const r = spawnSync(c, ['-version'], { encoding: 'utf8' })
+    if (r.status === 0) return c
+  }
+  return null
+}
+
 function run(cmd, args, opts = {}) {
   const r = spawnSync(cmd, args, { stdio: 'inherit', env: { ...process.env, ...(opts.env || {}) } })
   return r.status ?? 1
@@ -221,6 +237,44 @@ async function cmdAnalyze(args) {
   return run(py, argv)
 }
 
+/**
+ * Zero-dependency quick look: needs only ffmpeg, no python, no models.
+ * Produces a mel-ish spectrogram and a waveform picture.
+ */
+async function cmdSpectrogram(args) {
+  const audio = args._[0]
+  if (!audio) die('usage: music-lens spectrogram <audio> [--out DIR] [--size 1600x900]')
+  const src = path.resolve(audio)
+  if (!fs.existsSync(src)) die(`audio not found: ${src}`)
+
+  const ff = findFfmpeg()
+  if (!ff) {
+    die('ffmpeg not found.\n  install:  winget install Gyan.FFmpeg\n  or set:   MUSIC_LENS_FFMPEG=C:\\path\\to\\ffmpeg.exe')
+  }
+
+  const outDir = path.resolve(args.out || path.join(path.dirname(src), path.parse(src).name + '_look'))
+  fs.mkdirSync(outDir, { recursive: true })
+  const size = args.size || '1600x900'
+
+  log(`ffmpeg  : ${ff}`)
+  log(`output  : ${outDir}`)
+
+  const spec = path.join(outDir, 'spectrum.png')
+  const wave = path.join(outDir, 'waveform.png')
+
+  const st1 = run(ff, ['-y', '-hide_banner', '-loglevel', 'error', '-i', src,
+    '-lavfi', `showspectrumpic=s=${size}:legend=1:scale=log`, '-frames:v', '1', spec])
+  log(`  spectrum: ${st1 === 0 ? spec : 'FAILED'}`)
+
+  const st2 = run(ff, ['-y', '-hide_banner', '-loglevel', 'error', '-i', src,
+    '-lavfi', `showwavespic=s=${size}:split_channels=0:colors=#4a7ebb`, '-frames:v', '1', wave])
+  log(`  waveform: ${st2 === 0 ? wave : 'FAILED'}`)
+
+  if (st1 !== 0 && st2 !== 0) return 1
+  log('\nquick look done. for lyrics / key / tempo / numbers, run:  music-lens analyze "<same file>"')
+  return 0
+}
+
 async function cmdSeparate(args) {
   const audio = args._[0]
   if (!audio) die('usage: music-lens separate <audio> [--out DIR]')
@@ -242,6 +296,8 @@ function usage() {
               [--model medium]             small | medium | large-v3 | path
               [--lang auto]                language code or auto
               [--out DIR]                  output directory
+  music-lens spectrogram <audio>           quick look: spectrum + waveform (ffmpeg only,
+              [--size 1600x900]            no python, no models, ~10 seconds)
   music-lens separate <audio>              split vocals / instrumental (demucs)
 
 Outputs per song: 00_summary.md, 01_lyrics.txt, 02_transcript.json,
@@ -257,6 +313,7 @@ const map = {
   setup: cmdSetup,
   analyze: cmdAnalyze,
   separate: cmdSeparate,
+  spectrogram: cmdSpectrogram,
  }
 
 if (cmd === 'help' || cmd === '--help' || cmd === '-h') { usage(); process.exit(0) }

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "music-lens",
-  "version": "0.1.2",
+  "version": "0.1.4",
   "description": "Let a text-only model read a song: lyrics + timestamps, loudness, brightness, key, tempo, sections. Local, no API keys.",
   "type": "module",
   "bin": {

package/scripts/analyze.py

@@ -16,6 +16,7 @@ Usage:
 import argparse
 import json
 import os
+import re
 import sys
 import warnings
 from pathlib import Path
@@ -47,6 +48,35 @@ MINOR = np.array([6.33, 2.68, 3.52, 5.38, 2.60, 3.53, 2.54, 4.75, 3.98, 2.69, 3.
 NOTES = ["C", "C#", "D", "D#", "E", "F", "F#", "G", "G#", "A", "A#", "B"]
 NOTES_ZH = ["C", "升C", "D", "升D", "E", "F", "升F", "G", "升G", "A", "升A", "B"]
 
+# Whisper famously fabricates credits/subtitles when the audio carries no speech
+# (intros, outros, instrumental breaks). These are stock phrases from its training
+# data -- e.g. "作词 李宗盛" is a near-universal Chinese lyric-video credit line.
+HALLUCINATION_PATTERNS = [
+    r"作\s*[詞词]\s*[:：]?\s*\S{2,4}\s*作曲",     # "作词 X 作曲 Y"
+    r"作词曲",                                     # "作词曲 李宗盛"
+    r"[詞词曲]\s*[:：]\s*李宗盛",
+    r"李宗盛",
+    r"字幕\s*(by|BY|由|组|組|提供|制作|製作)",
+    r"中文字幕",
+    r"字幕组",
+    r"优优独播剧场",
+    r"YoYo Television",
+    r"Zither Harp",
+    r"请不吝点赞",
+    r"点赞.{0,4}订阅",
+    r"明镜与点点",
+    r"^\s*谢谢(观看|收看|收听)",
+    r"^\s*Thanks? for watching",
+    r"^\s*Subtitles?\s*by",
+    r"^\s*Amara\.org",
+]
+
+_HALLUCINATION_RE = re.compile("|".join(HALLUCINATION_PATTERNS), re.IGNORECASE)
+
+
+def is_hallucination(text: str) -> bool:
+    return bool(_HALLUCINATION_RE.search(text))
+
 
 def estimate_key(chroma_mean):
     best = (-2.0, None)
@@ -68,6 +98,8 @@ def main():
     ap.add_argument("--device", default="cpu")
     ap.add_argument("--skip-whisper", action="store_true")
     ap.add_argument("--vad", action="store_true", help="enable VAD (good for speech, bad for singing)")
+    ap.add_argument("--keep-hallucinations", action="store_true",
+                    help="keep fabricated credit/subtitle lines instead of dropping them")
     args = ap.parse_args()
 
     src = Path(args.audio).expanduser().resolve()
@@ -131,6 +163,7 @@ def main():
     # ------------------------------------------------------------- transcription
     segments = []
     info = {}
+    dropped = []
     if not args.skip_whisper:
         print(f"[2/6] transcribing with faster-whisper '{args.model}' (lang={args.lang}) ...")
         from faster_whisper import WhisperModel
@@ -170,6 +203,21 @@ def main():
         if len(deduped) != len(segments):
             print(f"      collapsed duplicates: {len(segments)} -> {len(deduped)} lines")
         segments = deduped
+
+        # drop fabricated credits / subtitles (whisper filler on instrumental parts)
+        dropped = []
+        if not args.keep_hallucinations:
+            kept = []
+            for seg in segments:
+                if is_hallucination(seg["text"]):
+                    dropped.append(seg)
+                    print(f"      dropped hallucination @{seg['start']:6.1f}s  {seg['text']}")
+                else:
+                    kept.append(seg)
+            if kept:
+                segments = kept
+            elif dropped:
+                print("      (all segments looked like hallucinations -- keeping them)")
     else:
         print("[2/6] whisper skipped")
 
@@ -253,6 +301,7 @@ def main():
         "vocal_time_ratio": round(sing_ratio, 2),
         "sections": sections,
         "segments_count": len(segments),
+        "dropped_hallucinations": len(dropped),
     }
     with open(out / "05_summary.json", "w", encoding="utf-8") as f:
         json.dump(summary, f, ensure_ascii=False, indent=2)
@@ -269,6 +318,9 @@ def main():
     L.append(f"- 动态范围：{lo_d} dB ～ {hi_d} dB（相对最响处）")
     L.append(f"- 人声占比：约 **{sing_ratio*100:.0f}%** 时长检测到人声")
     L.append(f"- 歌词字数：约 {total_lyrics} 字，{len(segments)} 句\n")
+    if dropped:
+        L.append(f"> 已过滤 {len(dropped)} 条疑似幻觉（片头/署名/字幕类）。"
+                 f"想保留请加 `--keep-hallucinations`。\n")
 
     L.append("## 段落结构（按响度自动切分）\n")
     for s in sections: