optimized PC player

media_player.py

@@ -35,6 +35,7 @@ import json
 import logging
 from logging.handlers import RotatingFileHandler
 import os
+import re
 import queue
 import sys
 import tempfile
@@ -86,6 +87,61 @@ DEBUG_LOG_DIR = Path.home() / ".mediaplayer_logs"
 DEBUG_LOG_PATH = DEBUG_LOG_DIR / "mediaplayer.log"
 _debug_logger = logging.getLogger("mediaplayer")
 
+_SPEAKER_TAG_RE = re.compile(r"\[(S\d+)\]")
+
+
+def _split_speaker_tagged_sentences(document: Dict[str, Any]) -> Dict[str, Any]:
+    """Split sentences that still contain inline ``[S01]`` speaker tags.
+
+    The server should already strip these tags and split at speaker boundaries
+    (see ``split_sentences_at_punctuation``), but locally cached boundaries
+    from an older server version may still have the tags embedded. This
+    post-processing ensures the player always uses clean, per-phrase sentences
+    for playback navigation, AI explanation, and every other operation.
+    """
+    sentences = document.get("sentences")
+    if not isinstance(sentences, list) or not sentences:
+        return document
+    needs_split = any(_SPEAKER_TAG_RE.search(str(s.get("text") or "")) for s in sentences)
+    if not needs_split:
+        return document
+
+    cleaned: list = []
+    for sentence in sentences:
+        text = str(sentence.get("text") or "")
+        matches = list(_SPEAKER_TAG_RE.finditer(text))
+        if not matches:
+            cleaned.append(sentence)
+            continue
+
+        start_ms = int(sentence["start_ms"])
+        end_ms = int(sentence["end_ms"])
+        duration_ms = max(1, end_ms - start_ms)
+        total = len(matches)
+        for index, match in enumerate(matches):
+            piece_start = match.end()
+            if index + 1 < total:
+                piece_end = matches[index + 1].start()
+            else:
+                piece_end = len(text)
+            piece = text[piece_start:piece_end].strip()
+            if not piece:
+                continue
+            sub_start = start_ms + int(duration_ms * index / total)
+            sub_end = start_ms + int(duration_ms * (index + 1) / total)
+            if sub_end <= sub_start:
+                sub_end = sub_start + 1
+            cleaned.append({
+                **sentence,
+                "start_ms": sub_start,
+                "end_ms": sub_end,
+                "text": piece,
+            })
+
+    for index, sentence in enumerate(cleaned):
+        sentence["index"] = index
+    return {**document, "sentences": cleaned}
+
 
 def setup_debug_logging() -> Path:
     """Write concise diagnostics to a small local rotating log."""
@@ -1103,8 +1159,12 @@ class MediaPlayerApp(tk.Tk):
             text="按句跳转",
             variable=self.sentence_mode,
             command=self._on_sentence_mode,
+            takefocus=0,
         )
         self.sentence_mode_btn.pack(side="left", padx=(16, 0))
+        # 点击 Checkbutton 后 tk 仍会把焦点给它；把空格键重定向到播放/暂停，
+        # 避免误改复选按钮状态。
+        self.sentence_mode_btn.bind("<space>", self._on_space)
         self.btn_generate_boundaries = ttk.Button(
             controls,
             text="生成句边界",
@@ -1172,6 +1232,8 @@ class MediaPlayerApp(tk.Tk):
             ("<period>", lambda _e: self.seek_sentence(1)),
             ("<e>", lambda _e: self.show_ai_explanation()),
             ("<E>", lambda _e: self.show_ai_explanation()),
+            ("<a>", lambda _e: self.show_ai_explanation()),
+            ("<A>", lambda _e: self.show_ai_explanation()),
             ("<f>", lambda _e: self.toggle_fullscreen()),
             ("<F>", lambda _e: self.toggle_fullscreen()),
             ("<Escape>", lambda _e: self.leave_fullscreen()),
@@ -1622,7 +1684,7 @@ class MediaPlayerApp(tk.Tk):
             document = self._load_cached_boundaries(media_hash)
             if document and document.get("sentences"):
                 self._media_hashes[path] = media_hash
-                self._boundaries[path] = document
+                self._boundaries[path] = _split_speaker_tagged_sentences(document)
                 count = len(document["sentences"])
                 message = f"已加载本地句边界缓存：共 {count} 句"
                 debug_log(
@@ -1714,6 +1776,9 @@ class MediaPlayerApp(tk.Tk):
         if (self._boundaries.get(path) or {}).get("sentences"):
             if not messagebox.askyesno("生成句边界", "当前文件已有句边界，确定要重新生成吗？"):
                 return
+            force = True
+        else:
+            force = False
         self.sentence_mode.set(False)
         self._save_settings()
         self._analysis_running.add(path)
@@ -1723,11 +1788,11 @@ class MediaPlayerApp(tk.Tk):
         self._update_sentence_controls()
         threading.Thread(
             target=self._analyze_sentences,
-            args=(path, token),
+            args=(path, token, force),
             daemon=True,
         ).start()
 
-    def _analyze_sentences(self, path, token):
+    def _analyze_sentences(self, path, token, force=False):
         error = ""
         document = None
         source = "generated"
@@ -1744,19 +1809,26 @@ class MediaPlayerApp(tk.Tk):
             )
             self._media_hashes[path] = media_hash
 
+            if force:
+                self._report_progress(18, "正在重新生成句边界…")
+            else:
                 self._report_progress(18, "正在检查线上是否已有句边界…")
                 remote_document = self._fetch_remote_boundaries(media_hash)
                 if remote_document and remote_document.get("sentences"):
                     document = remote_document
+                    document = _split_speaker_tagged_sentences(document)
                     source = "server"
                     self._save_cached_boundaries(media_hash, document)
-            else:
+            if document is None and not force:
                 document = self._load_cached_boundaries(media_hash)
                 if document and document.get("sentences"):
                     source = "local"
-                else:
+            if document is None:
+                if not force:
                     self._report_progress(25, "线上没有句边界，正在准备音频…")
-                    debug_log("audio_extraction_started", media_hash=media_hash)
+                else:
+                    self._report_progress(25, "正在重新提取音频…")
+                debug_log("audio_extraction_started", media_hash=media_hash, force=force)
                 audio_path = self._make_audio_upload(
                     path,
                     progress_callback=lambda ratio: self._report_progress(
@@ -1784,6 +1856,7 @@ class MediaPlayerApp(tk.Tk):
                         )
                     ),
                 )
+                document = _split_speaker_tagged_sentences(document)
                 self._save_cached_boundaries(media_hash, document)
                 source = "generated"
         except Exception as exc:
@@ -1800,6 +1873,7 @@ class MediaPlayerApp(tk.Tk):
 
         self._analysis_running.discard(path)
         if document is not None and document.get("sentences"):
+            document = _split_speaker_tagged_sentences(document)
             self._boundaries[path] = document
             count = len(document["sentences"])
             if source == "server":
@@ -2367,54 +2441,138 @@ class MediaPlayerApp(tk.Tk):
                 pass
         window = tk.Toplevel(self)
         window.title("AI 句子讲解")
-        window.geometry("480x380")
+        window.geometry("1000x800")
+        window.minsize(1000, 800)
+        window.configure(background="#1e1e2e")
         window.transient(self)
-        frame = ttk.Frame(window, padding=14)
+
+        # Dark theme palette.
+        bg = "#1e1e2e"
+        card_bg = "#2a2a3d"
+        fg_primary = "#cdd6f4"
+        fg_secondary = "#9399b2"
+        accent = "#89b4fa"
+        accent_hover = "#74c7ec"
+
+        frame = tk.Frame(window, bg=bg, padx=16, pady=14)
         frame.pack(fill="both", expand=True)
         frame.columnconfigure(0, weight=1)
-        frame.rowconfigure(1, weight=1)
+        frame.rowconfigure(2, weight=1)
-        ttk.Label(
+
-            frame,
+        font_size = [20]
-            text=sentence_text,
+
-            wraplength=440,
+        # ── Top toolbar panel ─────────────────────────────────────────────
-            justify="left",
+        # Fixed-height, always visible: title | A− 20pt A+ | ✕
-            foreground="#111111",
+        toolbar = tk.Frame(frame, bg=card_bg, padx=12, pady=8)
-            font=("Helvetica", 12, "bold"),
+        toolbar.grid(row=0, column=0, sticky="ew", pady=(0, 10))
-        ).grid(
+        toolbar.columnconfigure(1, weight=1)
-            row=0, column=0, sticky="ew", pady=(0, 10)
+
+        tk.Label(
+            toolbar,
+            text="AI 句子讲解",
+            bg=card_bg,
+            fg=fg_primary,
+            font=("Helvetica", 13, "bold"),
+        ).grid(row=0, column=0, sticky="w")
+
+        font_controls = tk.Frame(toolbar, bg=card_bg)
+        font_controls.grid(row=0, column=2, sticky="e")
+
+        def _apply_font_size():
+            size_label.configure(text=f"{font_size[0]}pt")
+            sentence_label.configure(font=("Helvetica", font_size[0], "bold"))
+            text.configure(font=("Helvetica", font_size[0]))
+
+        def _set_font_size(delta):
+            font_size[0] = max(12, min(36, font_size[0] + delta))
+            _apply_font_size()
+
+        tk.Button(
+            font_controls, text="A\u2212", command=lambda: _set_font_size(-2),
+            bg=accent, fg=bg, activebackground=accent_hover, activeforeground=bg,
+            relief="flat", padx=10, pady=4, font=("Helvetica", 12, "bold"),
+            takefocus=0, cursor="hand2",
+        ).pack(side="left", padx=(0, 4))
+
+        size_label = tk.Label(
+            font_controls, text=f"{font_size[0]}pt",
+            bg=card_bg, fg=fg_primary,
+            font=("Helvetica", 11), width=4, anchor="center",
         )
+        size_label.pack(side="left")
+
+        tk.Button(
+            font_controls, text="A+", command=lambda: _set_font_size(2),
+            bg=accent, fg=bg, activebackground=accent_hover, activeforeground=bg,
+            relief="flat", padx=10, pady=4, font=("Helvetica", 12, "bold"),
+            takefocus=0, cursor="hand2",
+        ).pack(side="left", padx=(4, 0))
+
+        tk.Button(
+            toolbar, text="\u2715",
+            command=lambda: self._close_explanation(generation),
+            bg=card_bg, fg=fg_secondary,
+            activebackground=accent_hover, activeforeground=bg,
+            relief="flat", padx=8, pady=4, font=("Helvetica", 13),
+            takefocus=0, cursor="hand2",
+        ).grid(row=0, column=3, sticky="e", padx=(12, 0))
+
+        # ── Sentence text (source) ────────────────────────────────────────
+        sentence_card = tk.Frame(frame, bg=card_bg, padx=12, pady=10)
+        sentence_card.grid(row=1, column=0, sticky="ew", pady=(0, 10))
+        sentence_card.columnconfigure(0, weight=1)
+        tk.Label(
+            sentence_card,
+            text="当前句子",
+            bg=card_bg,
+            fg=fg_secondary,
+            font=("Helvetica", 10),
+        ).grid(row=0, column=0, sticky="w")
+        sentence_label = tk.Label(
+            sentence_card,
+            text=sentence_text,
+            wraplength=940,
+            justify="left",
+            bg=card_bg,
+            fg=fg_primary,
+            font=("Helvetica", font_size[0], "bold"),
+        )
+        sentence_label.grid(row=1, column=0, sticky="ew", pady=(4, 0))
+
+        # ── Explanation text area ──────────────────────────────────────────
         text = tk.Text(
             frame,
             wrap="word",
             relief="flat",
-            background="#f6f6f6",
+            bg=card_bg,
-            foreground="#111111",
+            fg=fg_primary,
-            insertbackground="#111111",
+            insertbackground=fg_primary,
-            selectbackground="#3f8efc",
+            selectbackground=accent,
-            selectforeground="#ffffff",
+            selectforeground=bg,
-            padx=10,
+            padx=14,
-            pady=10,
+            pady=12,
+            font=("Helvetica", font_size[0]),
+            borderwidth=0,
+            highlightthickness=0,
         )
-        text.grid(row=1, column=0, sticky="nsew")
+        text.grid(row=2, column=0, sticky="nsew")
         text.insert("1.0", "AI 正在讲解这句话…")
         text.configure(state="disabled")
+
         self._explanation_text = text
         self._explanation_debug_var = tk.StringVar(
-            value=f"请求 #{generation}；连接 15 秒 / 响应最长 180 秒；日志见“调试日志”。"
+            value=f'请求 #{generation}；连接 15 秒 / 响应最长 180 秒；日志见\u201c调试日志\u201d。'
         )
         ttk.Label(
             frame,
             textvariable=self._explanation_debug_var,
-            wraplength=440,
+            wraplength=940,
             justify="left",
-            foreground="#666666",
+            foreground=fg_secondary,
-        ).grid(row=2, column=0, sticky="ew")
+            background=bg,
-        ttk.Button(
+        ).grid(row=3, column=0, sticky="ew", pady=(8, 0))
-            frame,
+
-            text="关闭",
+        window.bind("<Escape>", lambda _e: self._close_explanation(generation))
-            command=lambda: self._close_explanation(generation),
-            takefocus=0,
-        ).grid(row=3, column=0, sticky="e", pady=(10, 0))
         self._explanation_window = window
         window.protocol("WM_DELETE_WINDOW", lambda: self._close_explanation(generation))
 
@@ -2532,7 +2690,6 @@ class MediaPlayerApp(tk.Tk):
         try:
             text.configure(state="normal")
             text.delete("1.0", "end")
-            text.configure(foreground="#111111")
             text.insert("1.0", explanation)
             text.configure(state="disabled")
             if error:

moss/README.md

@@ -61,6 +61,7 @@ nvidia-smi
 ```bash
 sudo apt update
 sudo apt install -y ffmpeg git curl
+sudo apt install -y ninja-build
 ```
 
 ### 安装 uv
@@ -89,8 +90,20 @@ source .venv/bin/activate
 ```bash
 uv pip install -e ".[torch-runtime]" --torch-backend=auto
 uv pip install sglang-omni
+uv pip install ninja
 ```
 
+> `ninja` 是 FlashInfer JIT 编译 CUDA kernel 所需的构建工具，缺少它会在启动时
+> 报 `FileNotFoundError: 'ninja'`。
+>
+> 必须用 `apt install ninja-build` 安装系统级 ninja（`/usr/bin/ninja`）。
+> `uv pip install ninja` 只安装到虚拟环境 `.venv/bin/`，systemd 服务的 PATH
+> 不包含它。如果只能用 venv 里的 ninja，在 systemd 服务的 `[Service]` 段加：
+>
+> ```ini
+> Environment=PATH=/opt/moss-transcribe/MOSS-Transcribe-Diarize/.venv/bin:/usr/local/sbin:/usr/local/bin:/usr/sbin:/usr/bin:/sbin:/bin
+> ```
+
 ### 下载模型
 
 ```bash
@@ -111,8 +124,6 @@ sgl-omni serve \
   --model-path OpenMOSS-Team/MOSS-Transcribe-Diarize \
   --host 127.0.0.1 \
   --port 8001 \
-  --max-running-requests 4 \
-  --cuda-graph-max-bs 4 \
   --mem-fraction-static 0.35
 ```
 
@@ -122,9 +133,7 @@ sgl-omni serve \
 |------|--------|------|
 | `--host` | `127.0.0.1` | 只监听本机，不要暴露公网 |
 | `--port` | `8001` | 服务端口（Whisper 用 9000，避免冲突） |
-| `--max-running-requests` | `4` | 最大并发请求数；显存充足可调大 |
+| `--mem-fraction-static` | 自动 | 静态显存分配比例（作用于所有 SGLang engine stage）；与其他 GPU 服务共存时调低 |
-| `--cuda-graph-max-bs` | `4` | CUDA Graph 批大小上限，与并发数一致 |
-| `--mem-fraction-static` | `0.80` | 静态显存分配比例；与其他 GPU 服务共存时调低 |
 
 > 与 llama-server、Whisper 等其他 GPU 服务共存时，将
 > `--mem-fraction-static` 设为 `0.30`–`0.40`（约 7–10 GB），避免 OOM。
@@ -352,8 +361,7 @@ curl http://127.0.0.1:8000/healthz   # 确认 moss_configured 为 true
 | Speaches (Whisper) | 4 GB | 停掉后释放 |
 | **MOSS + llama-server** | ~16 GB / 24 GB | 安全 |
 
-独占 GPU 时可将 `--mem-fraction-static` 调到 `0.80`（约 19 GB），
+独占 GPU 时可将 `--mem-fraction-static` 调到 `0.80`（约 19 GB），支持更多并发。
-`--max-running-requests` 调到 `16`，支持更多并发。
 
 ---
 

moss/systemd/moss-transcribe-vllm.service

@@ -4,7 +4,7 @@ After=network-online.target
 
 [Service]
 Type=simple
-User=oraltrainer
+User=lsm
 WorkingDirectory=/opt/moss-transcribe
 Environment=HF_ENDPOINT=https://hf-mirror.com
 ExecStart=/opt/moss-transcribe/.venv/bin/vllm serve OpenMOSS-Team/MOSS-Transcribe-Diarize --host 127.0.0.1 --port 8001 --trust-remote-code

moss/systemd/moss-transcribe.service

@@ -4,15 +4,13 @@ After=network-online.target
 
 [Service]
 Type=simple
-User=oraltrainer
+User=lsm
 WorkingDirectory=/opt/moss-transcribe/MOSS-Transcribe-Diarize
 Environment=HF_ENDPOINT=https://hf-mirror.com
 ExecStart=/opt/moss-transcribe/MOSS-Transcribe-Diarize/.venv/bin/sgl-omni serve \
   --model-path OpenMOSS-Team/MOSS-Transcribe-Diarize \
   --host 127.0.0.1 \
   --port 8001 \
-  --max-running-requests 4 \
-  --cuda-graph-max-bs 4 \
   --mem-fraction-static 0.35
 Restart=on-failure
 RestartSec=10

sentence_api/transcription.py

@@ -246,6 +246,59 @@ _ABBREVIATION_RE = re.compile(
     r"e\.g\.|i\.e\.|no\.|approx\.|fig\.|inc\.|ltd\.|jr\.|sr\.|u\.s\.|u\.k\.)"
     r"(?![A-Za-z0-9])"
 )
+_SPEAKER_TAG_RE = re.compile(r"\[(S\d+)\]")
+
+
+def _expand_speaker_segments(
+    segment: TranscriptionSegment,
+) -> List[TranscriptionSegment]:
+    """Split a segment at inline ``[S01]`` speaker tags and strip the tags.
+
+    MOSS sometimes embeds speaker tags directly in the text (rather than in
+    the ``speaker`` JSON field). Each tag marks the start of a new speaker
+    turn, so the text is split there and the tags are removed from the
+    visible transcription.
+    """
+    text = segment.text
+    matches = list(_SPEAKER_TAG_RE.finditer(text))
+    if not matches:
+        return [segment]
+
+    sub_segments: List[TranscriptionSegment] = []
+    total = len(matches)
+    duration = segment.end_seconds - segment.start_seconds
+    for index, match in enumerate(matches):
+        start_pos = match.end()
+        if index + 1 < total:
+            end_pos = matches[index + 1].start()
+        else:
+            end_pos = len(text)
+        piece = text[start_pos:end_pos].strip()
+        if not piece:
+            continue
+        # Allocate time proportionally across sub-segments.
+        frac_start = index / total
+        frac_end = (index + 1) / total
+        sub_start = segment.start_seconds + duration * frac_start
+        sub_end = segment.start_seconds + duration * frac_end
+        words = None
+        if segment.words:
+            sub_words = [
+                w
+                for w in segment.words
+                if w.start_seconds >= sub_start and w.end_seconds <= sub_end
+            ]
+            words = sub_words or None
+        sub_segments.append(
+            TranscriptionSegment(
+                start_seconds=sub_start,
+                end_seconds=sub_end,
+                text=piece,
+                speaker=match.group(1),
+                words=words,
+            )
+        )
+    return sub_segments or [segment]
 
 
 def split_sentences_at_punctuation(
@@ -257,14 +310,18 @@ def split_sentences_at_punctuation(
     across segments and a sentence is closed at periods/question marks. Commas
     are also eligible boundaries, but only when the text before the comma has
     enough words; otherwise the comma stays with the following text.
+
+    Inline ``[S01]`` speaker tags are stripped and each speaker turn always
+    starts a new sentence, even when no sentence-ending punctuation exists.
     """
     sentences: List[TranscriptionSegment] = []
     pieces: List[str] = []
     start_seconds: Optional[float] = None
     last_segment_end: Optional[float] = None
+    current_speaker: Optional[str] = None
 
     def flush(end_seconds: float) -> None:
-        nonlocal pieces, start_seconds
+        nonlocal pieces, start_seconds, current_speaker
         text = " ".join(piece for piece in pieces if piece).strip()
         if text and start_seconds is not None and end_seconds > start_seconds:
             sentences.append(
@@ -272,19 +329,31 @@ def split_sentences_at_punctuation(
                     start_seconds=start_seconds,
                     end_seconds=end_seconds,
                     text=text,
+                    speaker=current_speaker,
                 )
             )
         pieces = []
         start_seconds = None
+        current_speaker = None
+
+    expanded: List[TranscriptionSegment] = []
+    for seg in segments:
+        expanded.extend(_expand_speaker_segments(seg))
 
     for segment in sorted(
-        segments, key=lambda item: (item.start_seconds, item.end_seconds)
+        expanded, key=lambda item: (item.start_seconds, item.end_seconds)
     ):
         text = segment.text.strip()
         if not text:
             continue
+        # Each speaker-tagged sub-segment always starts a new sentence,
+        # regardless of whether the speaker actually changed — MOSS emits
+        # a tag per phrase/line even for the same speaker.
+        if pieces and segment.speaker is not None:
+            flush(segment.start_seconds)
         if not pieces:
             start_seconds = segment.start_seconds
+        current_speaker = segment.speaker
         seg_pieces: List[Tuple[str, Optional[str]]] = []
         cursor = 0
         for match in _PUNCTUATION_BOUNDARY.finditer(text):
@@ -322,6 +391,7 @@ def split_sentences_at_punctuation(
                 end_seconds = ends[index] if index < len(ends) else segment.end_seconds
                 flush(end_seconds)
                 start_seconds = end_seconds
+                current_speaker = segment.speaker
     if pieces:
         flush(last_segment_end if last_segment_end is not None else start_seconds)
     return sentences