add sensevoice stt model

mlx_audio/stt/models/__init__.py

@@ -3,6 +3,7 @@
     lasr_ctc,
     parakeet,
     qwen3_asr,
+    sensevoice,
     voxtral,
     voxtral_realtime,
     wav2vec,

mlx_audio/stt/models/sensevoice/README.md

@@ -0,0 +1,39 @@
+# SenseVoice
+
+MLX implementation of [SenseVoice Small](https://github.com/FunAudioLLM/SenseVoice) from Alibaba DAMO Academy. Non-autoregressive CTC-based model with speech recognition, language identification, emotion recognition, and audio event detection.
+
+## Features
+
+- 50+ languages (strong on zh, en, ja, ko, yue)
+- emotion detection (happy, sad, angry, neutral, etc.)
+- audio event detection (speech, bgm, laughter, applause)
+- non-autoregressive inference (very fast, ~70ms for 10s audio)
+- ~234M parameters
+
+## Usage
+
+```python
+from mlx_audio.stt import load
+
+# downloads automatically from huggingface
+model = load("mlx-community/SenseVoiceSmall")
+
+result = model.generate("audio.wav", language="auto")
+print(result.text)
+print(result.language)
+
+# rich info (emotion, event) is in the first segment
+seg = result.segments[0]
+print(seg["emotion"])  # e.g. "neutral"
+print(seg["event"])    # e.g. "Speech"
+```
+
+### Language options
+
+`"auto"`, `"zh"`, `"en"`, `"ja"`, `"ko"`, `"yue"`, `"nospeech"`
+
+### Inverse text normalization
+
+```python
+result = model.generate("audio.wav", language="auto", use_itn=True)
+```

mlx_audio/stt/models/sensevoice/__init__.py

@@ -0,0 +1,3 @@
+from .config import ModelConfig
+from .sensevoice import SenseVoiceSmall
+from .sensevoice import SenseVoiceSmall as Model

mlx_audio/stt/models/sensevoice/config.py

@@ -0,0 +1,95 @@
+from dataclasses import dataclass, field
+from typing import Any, Dict, List, Optional
+
+
+@dataclass
+class EncoderConfig:
+    output_size: int = 512
+    attention_heads: int = 4
+    linear_units: int = 2048
+    num_blocks: int = 50
+    tp_blocks: int = 20
+    dropout_rate: float = 0.1
+    positional_dropout_rate: float = 0.1
+    attention_dropout_rate: float = 0.1
+    kernel_size: int = 11
+    sanm_shift: int = 0
+    normalize_before: bool = True
+
+    @classmethod
+    def from_dict(cls, params: Dict[str, Any]) -> "EncoderConfig":
+        import inspect
+
+        valid = {
+            k: v for k, v in params.items() if k in inspect.signature(cls).parameters
+        }
+        # Handle the typo in upstream config ("sanm_shfit" -> "sanm_shift")
+        if "sanm_shfit" in params and "sanm_shift" not in params:
+            valid["sanm_shift"] = params["sanm_shfit"]
+        return cls(**valid)
+
+
+@dataclass
+class FrontendConfig:
+    fs: int = 16000
+    window: str = "hamming"
+    n_mels: int = 80
+    frame_length: int = 25
+    frame_shift: int = 10
+    lfr_m: int = 7
+    lfr_n: int = 6
+
+    @classmethod
+    def from_dict(cls, params: Dict[str, Any]) -> "FrontendConfig":
+        import inspect
+
+        valid = {
+            k: v for k, v in params.items() if k in inspect.signature(cls).parameters
+        }
+        return cls(**valid)
+
+
+@dataclass
+class ModelConfig:
+    model_type: str = "sensevoice"
+    vocab_size: int = 25055
+    input_size: int = 560
+    encoder_conf: Optional[EncoderConfig] = None
+    frontend_conf: Optional[FrontendConfig] = None
+    # CMVN stats (loaded from am.mvn)
+    cmvn_means: Optional[List[float]] = None
+    cmvn_istd: Optional[List[float]] = None
+    # Path for loading ancillary files
+    model_path: Optional[str] = None
+
+    def __post_init__(self):
+        if self.encoder_conf is None:
+            self.encoder_conf = EncoderConfig()
+        elif isinstance(self.encoder_conf, dict):
+            self.encoder_conf = EncoderConfig.from_dict(self.encoder_conf)
+        if self.frontend_conf is None:
+            self.frontend_conf = FrontendConfig()
+        elif isinstance(self.frontend_conf, dict):
+            self.frontend_conf = FrontendConfig.from_dict(self.frontend_conf)
+
+    @classmethod
+    def from_dict(cls, params: Dict[str, Any]) -> "ModelConfig":
+        import inspect
+
+        encoder_conf = params.pop("encoder_conf", None)
+        frontend_conf = params.pop("frontend_conf", None)
+        valid = {
+            k: v for k, v in params.items() if k in inspect.signature(cls).parameters
+        }
+        config = cls(**valid)
+        if encoder_conf is not None:
+            if isinstance(encoder_conf, dict):
+                config.encoder_conf = EncoderConfig.from_dict(encoder_conf)
+            else:
+                config.encoder_conf = encoder_conf
+        if frontend_conf is not None:
+            if isinstance(frontend_conf, dict):
+                config.frontend_conf = FrontendConfig.from_dict(frontend_conf)
+            else:
+                config.frontend_conf = frontend_conf
+        return config