dsh-podwave 0.11.27 → 0.11.29

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
package/src/host.js CHANGED
@@ -56,6 +56,7 @@ import {
56
56
  deleteTranscript,
57
57
  getTranscriptionState,
58
58
  migrateAllTranscripts,
59
+ sweepWorkDirs,
59
60
  readTranscript,
60
61
  runtimeInfo,
61
62
  startTranscription,
@@ -68,10 +69,12 @@ import {
68
69
  editTranscriptLine,
69
70
  enqueueBatch,
70
71
  mergeTranscriptLine,
72
+ resegmentTranscript,
71
73
  saveTranscript,
72
74
  splitTranscriptLine,
73
75
  validateTranscript,
74
76
  } from './transcribe.js'
77
+ import { closeLiveSpeakers } from './live-speaker.js'
75
78
  import { isLinux, isMac, isWindows } from './platform.js'
76
79
  import { listJobs, removeJob, upsertJob, getJob } from './jobs.js'
77
80
  import { scan as scanSetup, plan as planSetup, install as installSetup, cancelSetup, setupJob, llamaBinary, verify as verifySetup } from './setup.js'
@@ -256,7 +259,8 @@ function boot() {
256
259
  booted = (async () => {
257
260
  await initState()
258
261
  loadXyz()
259
- await migrateAllTranscripts().catch((error) => console.error('[dsh-podwave] transcript migration failed:', error.message))
262
+ await sweepWorkDirs().catch(() => {})
263
+ await migrateAllTranscripts().catch((error) => console.error('[dsh-podwave] transcript migration failed:', error.message))
260
264
  if (getState().folders.length) {
261
265
  scanLibrary()
262
266
  .then(() => refreshWatchers())
@@ -490,7 +494,8 @@ async function route(ctx, req, res) {
490
494
  }
491
495
  if (routePath === 'api/setup/plan' && method === 'GET') {
492
496
  const source = url.searchParams.get('source') || 'auto'
493
- return json(res, 200, planSetup({ source }))
497
+ const mode = url.searchParams.get('mode') === 'local' ? 'local' : 'basic'
498
+ return json(res, 200, planSetup({ source, mode }))
494
499
  }
495
500
 
496
501
  if (routePath === 'api/setup/verify' && method === 'GET') {
@@ -509,7 +514,7 @@ async function route(ctx, req, res) {
509
514
  return json(res, 200, {
510
515
  ok: true,
511
516
  generatedAt: new Date().toISOString(),
512
- version: '0.11.27',
517
+ version: '0.11.29',
513
518
  platform,
514
519
  engine: recommendEngine(platform),
515
520
  paths: {
@@ -526,7 +531,8 @@ async function route(ctx, req, res) {
526
531
  if (routePath === 'api/setup/install' && method === 'POST') {
527
532
  const body = await readJson(req)
528
533
  if (setupJob()?.running) return json(res, 409, { error: '安装已在进行中' })
529
- installSetup({ source: body?.source || 'auto' }).catch(() => {})
534
+ // basic:只装 ffmpeg,配合云端 API 用;local:额外装 python 运行时 + 本地模型。
535
+ installSetup({ source: body?.source || 'auto', mode: body?.mode === 'local' ? 'local' : 'basic' }).catch(() => {})
530
536
  return json(res, 200, { started: true })
531
537
  }
532
538
  if (routePath === 'api/setup/cancel' && method === 'POST') {
@@ -538,7 +544,7 @@ async function route(ctx, req, res) {
538
544
  }
539
545
  if ((method === 'GET' || method === 'HEAD') && routePath === 'api/health') {
540
546
  await mediaReady
541
- return json(res, 200, { ok: true, app: 'podwave', version: '0.11.27', mediaBase, transcribe: runtimeInfo(), diarize: diarizeInfo(), translate: publicConfig() })
547
+ return json(res, 200, { ok: true, app: 'podwave', version: '0.11.29', mediaBase, transcribe: runtimeInfo(), diarize: diarizeInfo(), translate: publicConfig() })
542
548
  }
543
549
  if (method === 'GET' && routePath === 'api/state') {
544
550
  const state = getState()
@@ -921,7 +927,7 @@ async function route(ctx, req, res) {
921
927
  }
922
928
  if (!episode) return json(res, 404, { error: '单集不存在' })
923
929
  const body = await readJson(req)
924
- const result = await startTranscription(episode, { force: Boolean(body?.force), resume: Boolean(body?.resume) })
930
+ const result = await startTranscription(episode, { force: Boolean(body?.force), resume: Boolean(body?.resume), realtime: Boolean(body?.realtime) })
925
931
  return json(res, 200, result)
926
932
  }
927
933
 
@@ -959,15 +965,22 @@ async function route(ctx, req, res) {
959
965
  }
960
966
 
961
967
  if (routePath === 'api/transcribe/settings') {
962
- if (method === 'GET') return json(res, 200, { ...getSection('transcribe'), runtime: runtimeInfo() })
968
+ if (method === 'GET') {
969
+ // 密钥统一放在「DeepSeek API」那张卡片里,这里只回引擎选择;前端看 runtime.api.hasKey 判断填没填。
970
+ const section = getSection('transcribe') || {}
971
+ return json(res, 200, { ...section, runtime: runtimeInfo() })
972
+ }
963
973
  if (method === 'PUT') {
964
974
  const body = await readJson(req)
965
975
  const patch = {}
976
+ if (typeof body?.provider === 'string') patch.provider = body.provider === 'local' ? 'local' : 'api'
966
977
  if (Array.isArray(body?.hotwords)) patch.hotwords = body.hotwords.map((word) => String(word || '').trim()).filter(Boolean).slice(0, 500)
967
978
  if (Number.isFinite(Number(body?.chunkMinutes))) patch.chunkMinutes = Math.max(1, Math.min(60, Number(body.chunkMinutes)))
968
979
  if (typeof body?.diarize === 'boolean') patch.diarize = body.diarize
980
+ if (typeof body?.realtime === 'boolean') patch.realtime = body.realtime
969
981
  if (!Object.keys(patch).length) return json(res, 400, { error: '没有可保存的设置' })
970
- return json(res, 200, { ...updateSection('transcribe', patch), runtime: runtimeInfo() })
982
+ const saved = updateSection('transcribe', patch)
983
+ return json(res, 200, { ...saved, runtime: runtimeInfo() })
971
984
  }
972
985
  }
973
986
 
@@ -1018,6 +1031,21 @@ async function route(ctx, req, res) {
1018
1031
  }
1019
1032
  }
1020
1033
 
1034
+ const resegmentMatch = /^api\/episodes\/([^/]+)\/transcript\/resegment$/.exec(routePath)
1035
+ if (resegmentMatch && method === 'POST') {
1036
+ const id = decodeURIComponent(resegmentMatch[1])
1037
+ if (!findEpisode(id) && !isXyzId(id)) return json(res, 404, { error: '单集不存在' })
1038
+ const before = await readTranscript(id)
1039
+ if (!before) return json(res, 404, { error: '还没有转写' })
1040
+ try {
1041
+ pushUndo(id, before)
1042
+ const result = await resegmentTranscript(id)
1043
+ return json(res, 200, result)
1044
+ } catch (error) {
1045
+ return json(res, 400, { error: error.message })
1046
+ }
1047
+ }
1048
+
1021
1049
  const undoMatch = /^api\/episodes\/([^/]+)\/transcript\/undo$/.exec(routePath)
1022
1050
  if (undoMatch && method === 'POST') {
1023
1051
  const id = decodeURIComponent(undoMatch[1])
@@ -1150,7 +1178,25 @@ async function route(ctx, req, res) {
1150
1178
  })
1151
1179
  }
1152
1180
  items.sort((a, b) => (b.completedAt || 0) - (a.completedAt || 0))
1153
- return json(res, 200, { items })
1181
+ let bytes = 0
1182
+ try {
1183
+ for (const file of files) bytes += (await stat(path.join(TRANSCRIPTS_DIR, file))).size
1184
+ } catch {
1185
+ bytes = bytes || 0
1186
+ }
1187
+ return json(res, 200, { items, dir: TRANSCRIPTS_DIR, bytes })
1188
+ }
1189
+
1190
+ if (method === 'POST' && routePath === 'api/history/open') {
1191
+ try {
1192
+ fs.mkdirSync(TRANSCRIPTS_DIR, { recursive: true })
1193
+ const command = process.platform === 'darwin' ? 'open' : process.platform === 'win32' ? 'explorer' : 'xdg-open'
1194
+ const { spawn } = await import('node:child_process')
1195
+ spawn(command, [TRANSCRIPTS_DIR], { detached: true, stdio: 'ignore' }).unref()
1196
+ } catch {
1197
+ return json(res, 500, { error: '无法打开文件夹' })
1198
+ }
1199
+ return json(res, 200, { ok: true, dir: TRANSCRIPTS_DIR })
1154
1200
  }
1155
1201
 
1156
1202
  if (routePath === 'api/ui-settings') {
@@ -1160,6 +1206,16 @@ async function route(ctx, req, res) {
1160
1206
  const patch = {}
1161
1207
  if (typeof body?.showSidebarPanel === 'boolean') patch.showSidebarPanel = body.showSidebarPanel
1162
1208
  if (typeof body?.sidebarCollapsed === 'boolean') patch.sidebarCollapsed = body.sidebarCollapsed
1209
+ if (Number.isFinite(Number(body?.subtitleSize))) patch.subtitleSize = Math.max(12, Math.min(42, Math.round(Number(body.subtitleSize))))
1210
+ if (typeof body?.captionOpen === 'boolean') patch.captionOpen = body.captionOpen
1211
+ // 悬浮字幕的位置:传 null / 空串表示回到默认位置(注意 Number(null) 是 0,不能直接判有限)。
1212
+ const coordinate = (value) => {
1213
+ if (value === null || value === undefined || value === '') return null
1214
+ const number = Number(value)
1215
+ return Number.isFinite(number) ? Math.round(number) : null
1216
+ }
1217
+ if (body && 'captionX' in body) patch.captionX = coordinate(body.captionX)
1218
+ if (body && 'captionY' in body) patch.captionY = coordinate(body.captionY)
1163
1219
  if (!Object.keys(patch).length) return json(res, 400, { error: '没有可保存的设置' })
1164
1220
  return json(res, 200, saveUiSettings(patch))
1165
1221
  }
@@ -1285,6 +1341,9 @@ export default {
1285
1341
  }),
1286
1342
  'dsh-podwave: /podwave routes',
1287
1343
  )
1288
- ctx.effect(() => () => stopAll(), 'dsh-podwave: stop transcription workers')
1344
+ ctx.effect(() => () => {
1345
+ stopAll()
1346
+ closeLiveSpeakers()
1347
+ }, 'dsh-podwave: stop transcription workers')
1289
1348
  },
1290
1349
  }
package/src/library.js CHANGED
@@ -42,7 +42,7 @@ function isInside(parent, child) {
42
42
  return rel !== '' && !rel.startsWith('..') && !path.isAbsolute(rel)
43
43
  }
44
44
 
45
- async function walk(dir, depth, files) {
45
+ async function walk(dir, depth, files, root, realRoot = root) {
46
46
  let entries
47
47
  try {
48
48
  entries = await fsp.readdir(dir, { withFileTypes: true })
@@ -53,14 +53,15 @@ async function walk(dir, depth, files) {
53
53
  if (entry.name.startsWith('.')) continue
54
54
  const full = path.join(dir, entry.name)
55
55
  if (entry.isDirectory()) {
56
- if (depth < SCAN_DEPTH) await walk(full, depth + 1, files)
56
+ if (depth < SCAN_DEPTH) await walk(full, depth + 1, files, root, realRoot)
57
57
  } else if (entry.isFile() && AUDIO_EXTENSIONS.has(path.extname(entry.name).toLowerCase())) {
58
- files.push(full)
58
+ // 记下这个文件属于哪个「添加进来的根文件夹」,前端才能按目录逐层浏览。
59
+ files.push({ file: full, root, realRoot })
59
60
  }
60
61
  }
61
62
  }
62
63
 
63
- async function readEpisode(file) {
64
+ async function readEpisode(file, root, realRoot = root) {
64
65
  const stat = await fsp.stat(file)
65
66
  const id = episodeId(file)
66
67
  const cacheKey = `${stat.mtimeMs}:${stat.size}`
@@ -107,6 +108,9 @@ async function readEpisode(file) {
107
108
  duration,
108
109
  hasCover,
109
110
  folder: path.dirname(file),
111
+ root,
112
+ // 相对根文件夹的路径('' 表示就在根目录里),前端靠它逐层浏览目录树。
113
+ rel: path.relative(realRoot || root || '', path.dirname(file)).split(path.sep).join('/'),
110
114
  file,
111
115
  }
112
116
  metaCache.set(id, { key: cacheKey, data })
@@ -125,31 +129,37 @@ async function doScan() {
125
129
  } catch {
126
130
  continue
127
131
  }
128
- await walk(folder, 0, files)
132
+ let realRoot = folder
133
+ try {
134
+ realRoot = await fsp.realpath(folder)
135
+ } catch {
136
+ realRoot = folder
137
+ }
138
+ await walk(folder, 0, files, folder, realRoot)
129
139
  }
130
140
 
131
141
  const unique = []
132
142
  const seen = new Set()
133
- for (const file of files) {
134
- let real = file
143
+ for (const entry of files) {
144
+ let real = entry.file
135
145
  try {
136
- real = await fsp.realpath(file)
146
+ real = await fsp.realpath(entry.file)
137
147
  } catch {
138
148
  /* keep the original path */
139
149
  }
140
150
  if (seen.has(real)) continue
141
151
  seen.add(real)
142
- unique.push(real)
152
+ unique.push({ file: real, root: entry.root, realRoot: entry.realRoot })
143
153
  }
144
154
 
145
155
  const items = []
146
156
  let cursor = 0
147
157
  async function worker() {
148
158
  while (cursor < unique.length) {
149
- const file = unique[cursor]
159
+ const entry = unique[cursor]
150
160
  cursor += 1
151
161
  try {
152
- items.push(await readEpisode(file))
162
+ items.push(await readEpisode(entry.file, entry.root, entry.realRoot))
153
163
  } catch {
154
164
  /* skip unreadable files */
155
165
  }
@@ -196,10 +206,10 @@ export function getEpisodePaths() {
196
206
  }
197
207
 
198
208
  export function getEpisodes() {
199
- return episodes.map(({ file, folder, ...rest }) => ({
209
+ return episodes.map(({ file, ...rest }) => ({
200
210
  ...rest,
201
- // Only expose the last two path segments to the client.
202
- location: folder.split(path.sep).slice(-2).join('/'),
211
+ // folder/root 是绝对路径,前端靠它们做逐层目录浏览;location 只是给人看的短路径。
212
+ location: (rest.folder || '').split(path.sep).slice(-2).join('/'),
203
213
  }))
204
214
  }
205
215
 
@@ -0,0 +1,242 @@
1
+ // 实时说话人:常驻声纹服务 + 在线聚类。
2
+ //
3
+ // 整段说话人分离(diarize.js)要等全部音频转完才跑,实时跟播用不了。
4
+ // 这里的做法是「边播边猜」:每行字幕出来就取那一小段音频算一个声纹向量,
5
+ // 和已知说话人的质心比余弦相似度,够像就归入,不像就新建一个。
6
+ // 播完以后 transcribe.js 还会用整段 diarize 再校正一遍。
7
+ //
8
+ // 声纹模型和 diarize 用的是同一份 3D-Speaker embedding 模型(本地模型目录里已有)。
9
+ import fs from 'node:fs'
10
+ import path from 'node:path'
11
+ import { spawn } from 'node:child_process'
12
+ import { fileURLToPath } from 'node:url'
13
+ import { venvPython } from './platform.js'
14
+ import { diarizationPaths } from './diarize.js'
15
+
16
+ const SCRIPT = path.join(path.dirname(fileURLToPath(import.meta.url)), '..', 'scripts', 'speaker_server.py')
17
+ const READY_TIMEOUT_MS = 20000
18
+ const REQUEST_TIMEOUT_MS = 12000
19
+ // 同一个人在相邻片段上的余弦相似度一般 0.75 以上,不同人通常低于 0.5。
20
+ // 取 0.72 宁可多认出一个人,也别把两个人并成一个(后面还有整段校正兜底)。
21
+ const MATCH_THRESHOLD = 0.72
22
+ const MAX_SPEAKERS = 12
23
+
24
+ export function liveSpeakerInfo() {
25
+ const python = venvPython()
26
+ const { embModel } = diarizationPaths()
27
+ return {
28
+ available: fs.existsSync(python) && fs.existsSync(SCRIPT) && fs.existsSync(embModel),
29
+ python,
30
+ script: SCRIPT,
31
+ embModel,
32
+ }
33
+ }
34
+
35
+ function normalize(vector) {
36
+ let sum = 0
37
+ for (const value of vector) sum += value * value
38
+ const length = Math.sqrt(sum) || 1
39
+ return vector.map((value) => value / length)
40
+ }
41
+
42
+ function cosine(left, right) {
43
+ let sum = 0
44
+ for (let index = 0; index < left.length; index += 1) sum += left[index] * right[index]
45
+ return sum
46
+ }
47
+
48
+ // 常驻进程:模型加载一次要一两秒,每行都起进程肯定来不及。
49
+ class SpeakerServer {
50
+ constructor(child) {
51
+ this.child = child
52
+ this.pending = new Map()
53
+ this.nextId = 0
54
+ this.buffer = ''
55
+ this.closed = false
56
+ }
57
+
58
+ static async start(info) {
59
+ const child = spawn(info.python, [SCRIPT, '--model', info.embModel], {
60
+ stdio: ['pipe', 'pipe', 'pipe'],
61
+ env: { ...process.env, PYTHONUNBUFFERED: '1' },
62
+ })
63
+ const server = new SpeakerServer(child)
64
+ const ready = new Promise((resolve, reject) => {
65
+ const timer = setTimeout(() => reject(new Error('声纹服务启动超时')), READY_TIMEOUT_MS)
66
+ const onExit = () => { clearTimeout(timer); reject(new Error('声纹服务没能启动')) }
67
+ child.once('exit', onExit)
68
+ server.readyResolve = (value) => { clearTimeout(timer); child.off('exit', onExit); resolve(value) }
69
+ server.readyReject = (error) => { clearTimeout(timer); child.off('exit', onExit); reject(error) }
70
+ })
71
+ server.wire()
72
+ await ready
73
+ return server
74
+ }
75
+
76
+ wire() {
77
+ this.child.stdout.on('data', (chunk) => {
78
+ this.buffer += chunk.toString()
79
+ let index = this.buffer.indexOf('\n')
80
+ while (index >= 0) {
81
+ const line = this.buffer.slice(0, index).trim()
82
+ this.buffer = this.buffer.slice(index + 1)
83
+ if (line) this.handle(line)
84
+ index = this.buffer.indexOf('\n')
85
+ }
86
+ })
87
+ this.child.stderr.on('data', () => { /* 服务端日志忽略,出错会从 stdout 回来 */ })
88
+ this.child.on('exit', () => {
89
+ this.closed = true
90
+ for (const [, entry] of this.pending) entry.reject(new Error('声纹服务已退出'))
91
+ this.pending.clear()
92
+ if (this.readyReject) this.readyReject(new Error('声纹服务已退出'))
93
+ })
94
+ }
95
+
96
+ handle(line) {
97
+ let payload
98
+ try {
99
+ payload = JSON.parse(line)
100
+ } catch {
101
+ return
102
+ }
103
+ if (payload.ready) {
104
+ this.readyResolve?.(true)
105
+ return
106
+ }
107
+ if (payload.error && payload.id === undefined) {
108
+ this.readyReject?.(new Error(payload.error))
109
+ return
110
+ }
111
+ const entry = this.pending.get(payload.id)
112
+ if (!entry) return
113
+ this.pending.delete(payload.id)
114
+ clearTimeout(entry.timer)
115
+ if (payload.error) entry.reject(new Error(payload.error))
116
+ else entry.resolve(payload.embedding || null)
117
+ }
118
+
119
+ request(audio, start, end) {
120
+ if (this.closed) return Promise.reject(new Error('声纹服务已退出'))
121
+ const id = (this.nextId += 1)
122
+ return new Promise((resolve, reject) => {
123
+ const timer = setTimeout(() => {
124
+ this.pending.delete(id)
125
+ reject(new Error('声纹计算超时'))
126
+ }, REQUEST_TIMEOUT_MS)
127
+ this.pending.set(id, { resolve, reject, timer })
128
+ this.child.stdin.write(`${JSON.stringify({ id, audio, start, end })}\n`)
129
+ })
130
+ }
131
+
132
+ close() {
133
+ this.closed = true
134
+ try { this.child.stdin.end() } catch { /* 已经关了 */ }
135
+ try { this.child.kill() } catch { /* 已经退了 */ }
136
+ }
137
+ }
138
+
139
+ // 多个转写任务共用一个常驻进程(同一个 profile 只有一个用户)。
140
+ let sharedServer = null
141
+
142
+ async function ensureServer() {
143
+ const info = liveSpeakerInfo()
144
+ if (!info.available) return null
145
+ if (!sharedServer || sharedServer.closed) sharedServer = await SpeakerServer.start(info)
146
+ return sharedServer
147
+ }
148
+
149
+ async function sharedEmbed(audio, start, end) {
150
+ const server = await ensureServer()
151
+ if (!server) return null
152
+ return server.request(audio, start, end)
153
+ }
154
+
155
+ export function closeLiveSpeakers() {
156
+ if (sharedServer) {
157
+ sharedServer.close()
158
+ sharedServer = null
159
+ }
160
+ }
161
+
162
+ export function partForTime(parts, time) {
163
+ for (const part of parts) {
164
+ if (time >= part.start - 0.001 && time < part.end) return part
165
+ }
166
+ return parts[parts.length - 1] || null
167
+ }
168
+
169
+ export class OnlineSpeakers {
170
+ constructor(embed) {
171
+ this.embed = embed
172
+ this.centroids = []
173
+ this.counts = []
174
+ this.cache = new Map()
175
+ }
176
+
177
+ assign(vector) {
178
+ const unit = normalize(vector)
179
+ let best = -1
180
+ let bestScore = -Infinity
181
+ for (let index = 0; index < this.centroids.length; index += 1) {
182
+ const score = cosine(unit, this.centroids[index])
183
+ if (score > bestScore) {
184
+ bestScore = score
185
+ best = index
186
+ }
187
+ }
188
+ if (best >= 0 && bestScore >= MATCH_THRESHOLD) {
189
+ const count = this.counts[best] + 1
190
+ const merged = this.centroids[best].map((value, index) => (value * this.counts[best] + unit[index]) / count)
191
+ this.centroids[best] = normalize(merged)
192
+ this.counts[best] = count
193
+ return best
194
+ }
195
+ if (this.centroids.length >= MAX_SPEAKERS) return best >= 0 ? best : null
196
+ this.centroids.push(unit)
197
+ this.counts.push(1)
198
+ return this.centroids.length - 1
199
+ }
200
+
201
+ speakers() {
202
+ return this.centroids.map((_, index) => ({ id: index, label: `说话人 ${index + 1}` }))
203
+ }
204
+
205
+ // 给还没有 speaker 的行打实时标签。已有的不覆盖(可能来自整段校正)。
206
+ async labelLines(lines, parts) {
207
+ for (const line of lines) {
208
+ if (line.speaker !== null && line.speaker !== undefined) continue
209
+ const key = `${Math.round(line.start * 100)}|${line.text}`
210
+ let index = this.cache.get(key)
211
+ if (index === undefined) {
212
+ const part = partForTime(parts, line.start)
213
+ if (!part) continue
214
+ const start = Math.max(0, line.start - part.start)
215
+ const end = Math.min(part.duration, line.end - part.start)
216
+ if (end - start < 0.4) continue
217
+ let vector = null
218
+ try {
219
+ vector = await this.embed(part.file, start, end)
220
+ } catch {
221
+ vector = null
222
+ }
223
+ if (!vector) continue
224
+ index = this.assign(vector)
225
+ if (index === null) continue
226
+ this.cache.set(key, index)
227
+ }
228
+ line.speaker = index
229
+ }
230
+ }
231
+ }
232
+
233
+ // 拿不到本地声纹服务(没装 python 运行时/模型)就返回 null,实时只做转写和翻译。
234
+ export async function createLiveSpeakers() {
235
+ try {
236
+ const server = await ensureServer()
237
+ if (!server) return null
238
+ return new OnlineSpeakers(sharedEmbed)
239
+ } catch {
240
+ return null
241
+ }
242
+ }