fix: load official SenseVoice ONNX models

README.en.md

@@ -52,6 +52,19 @@ uv sync --extra whisper --extra web
 
 Available extras: `whisper`, `sensevoice`, `volcengine`, `web`, `server`.
 
+#### Prepare the SenseVoice ONNX model
+
+When using SenseVoice, download [`iic/SenseVoiceSmall-onnx`](https://modelscope.cn/models/iic/SenseVoiceSmall-onnx) to a local directory. That ONNX repository does not include the required `chn_jpn_yue_eng_ko_spectok.bpe.model`; download the file with that name from [`iic/SenseVoiceSmall`](https://modelscope.cn/models/iic/SenseVoiceSmall) and place it in the same directory. The final directory must contain at least:
+
+```text
+config.yaml
+am.mvn
+chn_jpn_yue_eng_ko_spectok.bpe.model
+model_quant.onnx  # official quantized model; model.onnx is also supported
+```
+
+Point the SenseVoice model-directory prompt at this directory. bili2text automatically selects quantized or non-quantized loading from `model_quant.onnx` or `model.onnx`, and reports incomplete directories before loading the runtime.
+
 ### Set Up
 
 A setup wizard runs automatically the first time, or you can launch it manually:

README.md

@@ -59,6 +59,19 @@ 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。
 
+#### 准备 SenseVoice ONNX 模型
+
+选择 SenseVoice 时，先下载 [`iic/SenseVoiceSmall-onnx`](https://modelscope.cn/models/iic/SenseVoiceSmall-onnx) 到本地目录。该 ONNX 仓库缺少运行时必需的 `chn_jpn_yue_eng_ko_spectok.bpe.model`，还需要从 [`iic/SenseVoiceSmall`](https://modelscope.cn/models/iic/SenseVoiceSmall) 下载这个同名文件，并放进同一目录。最终目录至少应包含：
+
+```text
+config.yaml
+am.mvn
+chn_jpn_yue_eng_ko_spectok.bpe.model
+model_quant.onnx  # 官方量化模型；也支持非量化 model.onnx
+```
+
+配置向导中的 SenseVoice 模型目录应指向这个目录。bili2text 会根据 `model_quant.onnx` 或 `model.onnx` 自动选择量化模式，并在文件不完整时列出缺失项。
+
 ### 初始化配置
 
 第一次运行时会自动弹出配置向导，也可以手动运行：

src/b2t/transcribers/sensevoice_local.py

@@ -6,11 +6,16 @@
 from b2t.i18n import dependency_sync_guidance
 from b2t.transcribers.base import Transcriber
 
+SENSEVOICE_TOKENIZER = "chn_jpn_yue_eng_ko_spectok.bpe.model"
+SENSEVOICE_SUPPORT_FILES = ("config.yaml", "am.mvn", SENSEVOICE_TOKENIZER)
+
 
 class SenseVoiceSmallTranscriber(Transcriber):
     name = "sensevoice"
 
-    def __init__(self, *, model_dir: Path, language: str = "auto", use_itn: bool = True) -> None:
+    def __init__(
+        self, *, model_dir: Path, language: str = "auto", use_itn: bool = True
+    ) -> None:
         self.model_dir = model_dir
         self.language = language
         self.use_itn = use_itn
@@ -28,7 +33,9 @@ def transcribe(
             progress.running("transcribing", message="transcribing", indeterminate=True)
 
         try:
-            from funasr_onnx.utils.postprocess_utils import rich_transcription_postprocess
+            from funasr_onnx.utils.postprocess_utils import (
+                rich_transcription_postprocess,
+            )
         except ImportError as exc:
             raise RuntimeError(
                 "SenseVoice support is not installed. "
@@ -58,7 +65,11 @@ def _ensure_model(self) -> Any:
             return self._model
 
         if not self.model_dir.exists():
-            raise RuntimeError(f"SenseVoice model directory does not exist: {self.model_dir}")
+            raise RuntimeError(
+                f"SenseVoice model directory does not exist: {self.model_dir}"
+            )
+
+        quantize = _validate_model_dir(self.model_dir)
 
         try:
             from funasr_onnx import SenseVoiceSmall
@@ -68,11 +79,36 @@ def _ensure_model(self) -> Any:
                 f"{dependency_sync_guidance('en-US')}"
             ) from exc
 
-        self._model = SenseVoiceSmall(str(self.model_dir))
+        self._model = SenseVoiceSmall(str(self.model_dir), quantize=quantize)
         return self._model
 
 
 def _extract_text(item: object) -> str:
     if isinstance(item, dict):
         return str(item.get("text", ""))
     return str(item)
+
+
+def _validate_model_dir(model_dir: Path) -> bool:
+    missing = [
+        filename
+        for filename in SENSEVOICE_SUPPORT_FILES
+        if not (model_dir / filename).is_file()
+    ]
+    if missing:
+        missing_text = ", ".join(missing)
+        raise RuntimeError(
+            f"SenseVoice model directory is incomplete ({model_dir}). Missing: {missing_text}. "
+            "The ONNX repository does not include the SentencePiece tokenizer; copy "
+            f"{SENSEVOICE_TOKENIZER} from iic/SenseVoiceSmall into this directory."
+        )
+
+    if (model_dir / "model_quant.onnx").is_file():
+        return True
+    if (model_dir / "model.onnx").is_file():
+        return False
+
+    raise RuntimeError(
+        f"SenseVoice model directory is incomplete ({model_dir}). "
+        "Missing model_quant.onnx or model.onnx."
+    )

tests/test_sensevoice_local.py

@@ -0,0 +1,59 @@
+from __future__ import annotations
+
+import sys
+from pathlib import Path
+from types import SimpleNamespace
+
+import pytest
+
+from b2t.transcribers.sensevoice_local import SenseVoiceSmallTranscriber
+
+REQUIRED_SUPPORT_FILES = (
+    "config.yaml",
+    "am.mvn",
+    "chn_jpn_yue_eng_ko_spectok.bpe.model",
+)
+
+
+def _write_support_files(model_dir: Path) -> None:
+    for filename in REQUIRED_SUPPORT_FILES:
+        (model_dir / filename).write_text("test", encoding="utf-8")
+
+
+@pytest.mark.parametrize(
+    ("model_filename", "expected_quantize"),
+    (("model_quant.onnx", True), ("model.onnx", False)),
+)
+def test_sensevoice_selects_quantization_from_available_model(
+    tmp_path: Path,
+    monkeypatch: pytest.MonkeyPatch,
+    model_filename: str,
+    expected_quantize: bool,
+) -> None:
+    _write_support_files(tmp_path)
+    (tmp_path / model_filename).write_bytes(b"onnx")
+    calls: list[tuple[str, bool]] = []
+
+    def fake_model(model_dir: str, *, quantize: bool):
+        calls.append((model_dir, quantize))
+        return object()
+
+    monkeypatch.setitem(
+        sys.modules, "funasr_onnx", SimpleNamespace(SenseVoiceSmall=fake_model)
+    )
+
+    transcriber = SenseVoiceSmallTranscriber(model_dir=tmp_path)
+    transcriber._ensure_model()
+
+    assert calls == [(str(tmp_path), expected_quantize)]
+
+
+def test_sensevoice_reports_missing_tokenizer_before_model_load(tmp_path: Path) -> None:
+    (tmp_path / "model_quant.onnx").write_bytes(b"onnx")
+    (tmp_path / "config.yaml").write_text("test", encoding="utf-8")
+    (tmp_path / "am.mvn").write_text("test", encoding="utf-8")
+
+    transcriber = SenseVoiceSmallTranscriber(model_dir=tmp_path)
+
+    with pytest.raises(RuntimeError, match="chn_jpn_yue_eng_ko_spectok.bpe.model"):
+        transcriber._ensure_model()