feat: 新增小米 MiMo 云端转写引擎（切片并发 + 断点续跑）

CHANGELOG.en.md

@@ -8,6 +8,14 @@
 - added multiline batch submission to the web UI and API
 - added multiline batch processing to the desktop window
 
+### MiMo Transcription Engine
+
+- added Xiaomi MiMo (`mimo-v2.5-asr`) as a cloud engine, using OpenAI-compatible chat completions with `input_audio`
+- long audio is split into 60s chunks and transcribed concurrently; added `--workers` and the `mimo.workers` config key (default 4)
+- chunks retry with backoff so a single rate-limit response cannot discard a long job
+- successful chunks are cached under `.b2t/cache/mimo/` keyed by audio content, so a re-run after an interruption only transcribes what is missing
+- fixed a crash when indexing local-file transcriptions whose `download` metadata is `None`
+
 ## 2026-04-10 (v0.3.0)
 
 ### Version Cleanup

CHANGELOG.md

@@ -8,6 +8,14 @@
 - Web/API 支持多行输入与批量任务提交
 - 桌面窗口支持在输入框中粘贴多行来源并逐条处理
 
+### MiMo 转写引擎
+
+- 新增小米 MiMo（`mimo-v2.5-asr`）云端引擎，走 OpenAI 兼容的 chat completions + `input_audio` 协议
+- 长音频按 60 秒切片并发转写，新增 `--workers` 与配置项 `mimo.workers`（默认 4）
+- 切片失败自动退避重试，避免单次限流让整个长任务作废
+- 转写成功的切片按音频内容哈希缓存到 `.b2t/cache/mimo/`，中断后重跑只补缺失部分
+- 修复本地文件转写时 `download` 元数据为 `None` 导致的入库报错
+
 ## 2026-04-10 (v0.3.0)
 
 ### 版本整理

README.en.md

@@ -31,6 +31,7 @@ There's also a simple web UI and a desktop window for anyone who'd rather not us
 | **Whisper** | Local model | OpenAI's open-source speech recognition model. Runs offline, general-purpose |
 | **SenseVoice** | Local model | ONNX-based local model with strong Chinese recognition |
 | **Volcengine** | Cloud API | ByteDance's commercial ASR service, good for batch or service-oriented workloads |
+| **MiMo** | Cloud API | Xiaomi MiMo ASR (`mimo-v2.5-asr`) over an OpenAI-compatible API; long audio is chunked and transcribed in parallel |
 
 ## Quick Start
 
@@ -50,7 +51,7 @@ This only installs core dependencies. Transcription engines and extra features a
 uv sync --extra whisper --extra web
 ```
 
-Available extras: `whisper`, `sensevoice`, `volcengine`, `web`, `server`.
+Available extras: `whisper`, `sensevoice`, `volcengine`, `mimo`, `web`, `server`.
 
 ### Set Up
 
@@ -92,6 +93,23 @@ Or put one BV, URL, or local file path per line:
 uv run bili2text batch --file sources.txt
 ```
 
+### MiMo cloud transcription
+
+MiMo uses an OpenAI-compatible API, so install the extra and configure a key first:
+
+```bash
+uv sync --extra mimo
+uv run bili2text bootstrap   # pick mimo, paste the API key
+```
+
+Long audio is split into 60-second chunks and transcribed concurrently. `--workers` sets the concurrency (default 4):
+
+```bash
+uv run bili2text tx "BV1kfDTBXEfu" --provider mimo --workers 6
+```
+
+Pushing concurrency too high gets you rate-limited (HTTP 429). Individual chunks retry with backoff, and successful chunks are cached under `.b2t/cache/mimo/` keyed by audio content — so re-running an interrupted job **only pays for what is missing**.
+
 ## Commands
 
 | Command | Alias | What it does |

README.md

@@ -36,6 +36,7 @@
 | **Whisper** | 本地模型 | OpenAI 开源的语音识别模型，离线运行，通用性强 |
 | **SenseVoice** | 本地模型 | 阿里云开源本地语音识别模型，中文识别效果好 |
 | **火山引擎** | 云端 API | 字节跳动旗下的商用语音识别服务，识别很准很推荐 |
+| **MiMo** | 云端 API | 小米 MiMo 语音识别（`mimo-v2.5-asr`），OpenAI 兼容协议，长音频自动切片并发转写 |
 
 ## 快速开始
 
@@ -57,7 +58,7 @@ uv sync
 uv sync --extra whisper --extra web
 ```
 
-可选的 extras：`whisper`、`sensevoice`、`volcengine`、`web`、`server`。可以暂时不用安装，详看下方的初始化文档。
+可选的 extras：`whisper`、`sensevoice`、`volcengine`、`mimo`、`web`、`server`。可以暂时不用安装，详看下方的初始化文档。
 
 ### 初始化配置
 
@@ -99,6 +100,23 @@ uv run bili2text batch "BV1kfDTBXEfu" "https://www.bilibili.com/video/BV1xx411c7
 uv run bili2text batch --file sources.txt
 ```
 
+### MiMo 云端转写
+
+MiMo 走的是 OpenAI 兼容协议，需要先装 extra 并配好 API Key：
+
+```bash
+uv sync --extra mimo
+uv run bili2text bootstrap   # 选 mimo，填 API Key
+```
+
+长音频会自动按 60 秒切片并发转写，`--workers` 控制并发数（默认 4）：
+
+```bash
+uv run bili2text tx "BV1kfDTBXEfu" --provider mimo --workers 6
+```
+
+并发调太高会被服务端限流（HTTP 429）。单个切片失败会自动退避重试；转写成功的切片按内容哈希缓存在 `.b2t/cache/mimo/`，所以任务中断后**原样重跑只会补缺失的部分**，不会重复计费。
+
 ## 命令一览
 
 | 命令 | 缩写 | 说明 |

pyproject.toml

@@ -30,6 +30,9 @@ sensevoice = [
 volcengine = [
   "requests>=2.32.3",
 ]
+mimo = [
+  "openai>=1.40.0",
+]
 web = [
   "fastapi>=0.115.12",
   "jinja2>=3.1.6",

src/b2t/bootstrap.py

@@ -164,6 +164,11 @@ def run_bootstrap(*, settings: Settings, interactive: bool = True) -> AppConfig:
             "value": "volcengine",
             "enabled": "volcengine" in config.enabled_providers,
         },
+        {
+            "name": f"mimo       — {tr(lang, 'provider_mimo_short')}",
+            "value": "mimo",
+            "enabled": "mimo" in config.enabled_providers,
+        },
     ]
     selected_providers: list[str] = inquirer.checkbox(
         message=tr(lang, "bootstrap_providers_prompt"),
@@ -212,6 +217,8 @@ def run_bootstrap(*, settings: Settings, interactive: bool = True) -> AppConfig:
             _configure_sensevoice(config, lang)
         elif provider == "volcengine":
             _configure_volcengine(config, lang)
+        elif provider == "mimo":
+            _configure_mimo(config, lang)
 
     # ── 5. Pick default provider ─────────────────────────────
     console.print()
@@ -324,6 +331,25 @@ def _configure_volcengine(config: AppConfig, lang: str) -> None:
     ).execute()
 
 
+def _configure_mimo(config: AppConfig, lang: str) -> None:
+    config.mimo.api_key = inquirer.secret(
+        message=tr(lang, "bootstrap_mimo_api_key_prompt"),
+        default=config.mimo.api_key,
+    ).execute().strip()
+    config.mimo.base_url = inquirer.text(
+        message=tr(lang, "bootstrap_mimo_base_url_prompt"),
+        default=config.mimo.base_url,
+    ).execute().strip()
+    config.mimo.model_name = inquirer.text(
+        message=tr(lang, "bootstrap_mimo_model_prompt"),
+        default=config.mimo.model_name,
+    ).execute().strip()
+    config.mimo.language = inquirer.text(
+        message=tr(lang, "bootstrap_mimo_language_prompt"),
+        default=config.mimo.language,
+    ).execute().strip()
+
+
 def _show_next_steps(
     *,
     console: Console,

src/b2t/cli.py

@@ -49,10 +49,14 @@ def transcribe(
         prompt: str = typer.Option("", "--prompt", help=tr(language, "opt_prompt_help")),
         output: Path | None = typer.Option(None, "--output", help=tr(language, "opt_output_help")),
         workspace: Path | None = typer.Option(None, "--workspace", help=tr(language, "opt_workspace_help")),
+        workers: int | None = typer.Option(None, "--workers", help=tr(language, "opt_workers_help")),
     ) -> None:
         """Download or open media, then transcribe it with the selected provider."""
         try:
             settings, config = _load_runtime(workspace=workspace, provider=provider, model=model)
+            if workers is not None:
+                # only MiMo splits audio into chunks and runs them in parallel today
+                config.mimo.workers = max(1, workers)
             renderer = TqdmTaskRenderer(config.language)
             service = _build_task_service(
                 settings=settings,

src/b2t/config.py

@@ -18,6 +18,7 @@ class Settings:
     transcripts_edited_dir: Path
     metadata_dir: Path
     tasks_dir: Path
+    cache_dir: Path
     config_path: Path
     app_db_path: Path
 
@@ -33,6 +34,7 @@ def from_workspace(cls, workspace: Path | None = None) -> "Settings":
             transcripts_edited_dir=root / "transcripts" / "edited",
             metadata_dir=root / "metadata",
             tasks_dir=root / "tasks",
+            cache_dir=root / "cache",
             config_path=root / "config.json",
             app_db_path=root / "app.db",
         )
@@ -47,5 +49,6 @@ def ensure_directories(self) -> None:
             self.transcripts_edited_dir,
             self.metadata_dir,
             self.tasks_dir,
+            self.cache_dir,
         ):
             directory.mkdir(parents=True, exist_ok=True)

src/b2t/factory.py

@@ -43,6 +43,20 @@ def build_pipeline(
             model_name=selected_model or config.volcengine.model_name,
             use_itn=config.volcengine.use_itn,
         )
+    elif selected_provider == "mimo":
+        from b2t.transcribers.mimo import MimoTranscriber
+
+        # selected_model may carry whisper's default ("small"); only honor it when it
+        # actually looks like a MiMo model id.
+        chosen_model = selected_model if selected_model.startswith("mimo") else config.mimo.model_name
+        transcriber = MimoTranscriber(
+            api_key=config.mimo.api_key,
+            base_url=config.mimo.base_url,
+            model_name=chosen_model or "mimo-v2.5-asr",
+            language=config.mimo.language,
+            workers=config.mimo.workers,
+            cache_dir=settings.cache_dir,
+        )
     else:
         raise RuntimeError(f"Unsupported provider: {selected_provider}")
 

src/b2t/i18n.py

@@ -25,11 +25,12 @@
         "cmd_language_help": "切换界面语言（缩写: lang）。",
         "arg_source_help": "BV 号、Bilibili 链接或本地文件路径。",
         "arg_sources_help": "一条或多条输入；也可以用 --file 读取每行一个输入。",
-        "opt_provider_help": "转写引擎: whisper / sensevoice / volcengine。",
+        "opt_provider_help": "转写引擎: whisper / sensevoice / volcengine / mimo。",
         "opt_model_help": "模型名称。",
         "opt_prompt_help": "转写提示词（可选）。",
         "opt_output_help": "输出文件或目录。",
         "opt_workspace_help": "工作目录，默认 ./.b2t。",
+        "opt_workers_help": "MiMo 分片并发数，默认读配置（4）。调高更快，太高会被服务端限流。",
         "opt_host_help": "监听地址。",
         "opt_port_help": "监听端口。",
         "opt_language_help": "语言代码，如 zh-CN、en-US。",
@@ -79,6 +80,10 @@
         "bootstrap_volc_resource_prompt": "火山引擎 Resource ID",
         "bootstrap_volc_model_prompt": "火山引擎模型名称",
         "bootstrap_volc_itn_prompt": "启用逆文本正则化",
+        "bootstrap_mimo_api_key_prompt": "小米 MiMo API Key（Token Plan）",
+        "bootstrap_mimo_base_url_prompt": "小米 MiMo Base URL",
+        "bootstrap_mimo_model_prompt": "小米 MiMo 模型名称",
+        "bootstrap_mimo_language_prompt": "识别语言（如 zh、en，留空表示自动）",
         "bootstrap_saved": "配置已保存到 {path}",
         "bootstrap_auto_start": "还没有配置文件，来走一下初始化向导吧……",
         "bootstrap_finish": "搞定！随时可以用 `bili2text bootstrap` 重新配置，或 `bili2text lang <代码>` 切换语言。",
@@ -113,6 +118,7 @@
         "provider_whisper_short": "本地离线，不依赖云服务",
         "provider_sensevoice_short": "本地 ONNX 模型，中文效果好",
         "provider_volcengine_short": "火山引擎云端识别，需要凭据",
+        "provider_mimo_short": "小米 MiMo 云端 ASR，OpenAI 兼容协议",
         "feature_web_short": "浏览器界面，在网页上操作",
         "feature_server_short": "服务模式，适合 Docker / 局域网部署",
         "feature_window_short": "简单的桌面窗口",
@@ -123,6 +129,8 @@
         "provider_sensevoice_desc": "基于 ONNX 的本地模型，中文识别效果不错。需要模型文件和对应依赖。",
         "provider_volcengine_name": "火山引擎云端识别",
         "provider_volcengine_desc": "商用云端语音识别，适合服务化部署。需要配置火山引擎凭据。",
+        "provider_mimo_name": "小米 MiMo（云端）",
+        "provider_mimo_desc": "小米 MiMo 云端语音识别（mimo-v2.5-asr），走 OpenAI chat completions + input_audio 协议。长音频自动切片并发转写。需要配置 API Key。",
 
         # ── Language ─────────────────────────────────────────
         "language_updated": "语言已切换为: {language}",
@@ -215,11 +223,12 @@
         "cmd_language_help": "Switch the interface language (alias: lang).",
         "arg_source_help": "BV id, Bilibili URL, or a local media file.",
         "arg_sources_help": "One or more inputs; use --file to read one input per line.",
-        "opt_provider_help": "Transcription engine: whisper / sensevoice / volcengine.",
+        "opt_provider_help": "Transcription engine: whisper / sensevoice / volcengine / mimo.",
         "opt_model_help": "Model name.",
         "opt_prompt_help": "Optional transcription prompt.",
         "opt_output_help": "Output file or directory.",
         "opt_workspace_help": "Workspace root, defaults to ./.b2t.",
+        "opt_workers_help": "MiMo chunk concurrency; defaults to config (4). Higher is faster until the API rate-limits you.",
         "opt_host_help": "Bind address.",
         "opt_port_help": "Bind port.",
         "opt_language_help": "Language code, e.g. zh-CN or en-US.",
@@ -269,6 +278,10 @@
         "bootstrap_volc_resource_prompt": "Volcengine resource ID",
         "bootstrap_volc_model_prompt": "Volcengine model name",
         "bootstrap_volc_itn_prompt": "Turn on inverse text normalization",
+        "bootstrap_mimo_api_key_prompt": "Xiaomi MiMo API key (Token Plan)",
+        "bootstrap_mimo_base_url_prompt": "MiMo base URL",
+        "bootstrap_mimo_model_prompt": "MiMo model name",
+        "bootstrap_mimo_language_prompt": "Recognition language (e.g. zh, en — leave blank for auto)",
         "bootstrap_saved": "Config saved to {path}",
         "bootstrap_auto_start": "No config file found — let's get you set up...",
         "bootstrap_finish": "You're all set! Run `bili2text bootstrap` anytime to reconfigure, or `bili2text lang <code>` to switch languages.",
@@ -303,6 +316,7 @@
         "provider_whisper_short": "Runs locally, no cloud needed",
         "provider_sensevoice_short": "Local ONNX model, great for Chinese",
         "provider_volcengine_short": "Volcengine cloud ASR, needs credentials",
+        "provider_mimo_short": "Xiaomi MiMo cloud ASR, OpenAI-compatible",
         "feature_web_short": "Web interface — use in your browser",
         "feature_server_short": "Server mode for Docker / LAN",
         "feature_window_short": "Simple desktop window",
@@ -313,6 +327,8 @@
         "provider_sensevoice_desc": "ONNX-based local model with strong Chinese recognition. Needs the model files and dependencies.",
         "provider_volcengine_name": "Volcengine (cloud)",
         "provider_volcengine_desc": "Commercial cloud-based speech recognition. Good for service deployments. Requires Volcengine credentials.",
+        "provider_mimo_name": "Xiaomi MiMo (cloud)",
+        "provider_mimo_desc": "Xiaomi MiMo cloud ASR (mimo-v2.5-asr) over OpenAI chat completions with input_audio. Long audio is chunked and transcribed in parallel. Requires an API key.",
 
         # ── Language ─────────────────────────────────────────
         "language_updated": "Language switched to: {language}",

src/b2t/library.py

@@ -31,7 +31,7 @@ def register_transcript_result(self, result: TranscriptResult) -> int:
             source_input=result.source.raw_input,
             source_url=result.source.url,
             source_bv=result.source.bv,
-            title=result.metadata.get("download", {}).get("title") or result.source.display_name,
+            title=(result.metadata.get("download") or {}).get("title") or result.source.display_name,
             display_name=result.source.display_name,
             language=result.metadata.get("language"),
             engine=result.engine,

src/b2t/transcribers/__init__.py

@@ -1,11 +1,13 @@
 from b2t.transcribers.base import Transcriber
+from b2t.transcribers.mimo import MimoTranscriber
 from b2t.transcribers.sensevoice_local import SenseVoiceSmallTranscriber
 from b2t.transcribers.volcengine import VolcengineFlashTranscriber
 from b2t.transcribers.whisper_local import LocalWhisperTranscriber
 
 __all__ = [
     "Transcriber",
     "LocalWhisperTranscriber",
+    "MimoTranscriber",
     "SenseVoiceSmallTranscriber",
     "VolcengineFlashTranscriber",
 ]