feat: add Qwen3-ASR engine with timestamp support and unit tests

app/engines/stt/qwen3asr/__init__.py

@@ -0,0 +1,6 @@
+"""Qwen3-ASR STT Engine"""
+
+from app.engines.stt.qwen3asr.config import Qwen3ASRConfig
+from app.engines.stt.qwen3asr.engine import Qwen3ASREngine
+
+__all__ = ["Qwen3ASRConfig", "Qwen3ASREngine"]

app/engines/stt/qwen3asr/config.py

@@ -0,0 +1,67 @@
+"""Configuration for Qwen3-ASR STT Engine"""
+
+from pydantic import Field, field_validator
+
+from app.models.engine import EngineConfig
+
+
+class Qwen3ASRConfig(EngineConfig):
+    """
+    Configuration for Qwen3-ASR STT engine (vLLM backend)
+
+    Extends EngineConfig with Qwen3-ASR specific parameters.
+
+    Attributes:
+        model_name: HuggingFace model ID (Qwen/Qwen3-ASR-1.7B or Qwen/Qwen3-ASR-0.6B)
+        dtype: Torch dtype for model loading (bfloat16 recommended)
+        gpu_memory_utilization: Fraction of GPU memory to use for vLLM
+        max_inference_batch_size: Maximum batch size for inference
+        max_new_tokens: Maximum tokens to generate (increase for long audio)
+    """
+
+    @field_validator("model_name")
+    @classmethod
+    def validate_model(cls, v: str) -> str:
+        """Validate Qwen3-ASR model name"""
+        valid = [
+            "Qwen/Qwen3-ASR-1.7B",
+            "Qwen/Qwen3-ASR-0.6B",
+        ]
+        # Allow local paths
+        if v in valid or v.startswith("/") or v.startswith("./"):
+            return v
+        raise ValueError(f"Qwen3-ASR model must be one of {valid} or a local path")
+
+    # vLLM backend specific fields
+    gpu_memory_utilization: float = Field(
+        default=0.7,
+        ge=0.1,
+        le=0.95,
+        description="Fraction of GPU memory to use for vLLM",
+    )
+    max_inference_batch_size: int = Field(
+        default=32,
+        gt=0,
+        description="Maximum batch size for inference. -1 for unlimited.",
+    )
+    max_new_tokens: int = Field(
+        default=512,
+        gt=0,
+        description="Maximum tokens to generate. Increase for long audio.",
+    )
+    forced_aligner: str | None = Field(
+        default=None, description="Path or name of ForcedAligner model for timestamps."
+    )
+    unfixed_chunk_num: int = Field(
+        default=4, description="Number of unfixed chunks for streaming state"
+    )
+    unfixed_token_num: int = Field(
+        default=5, description="Number of unfixed tokens for streaming state"
+    )
+    chunk_size_sec: float = Field(
+        default=1.0, description="Size of each audio chunk in seconds for streaming"
+    )
+    language: str | None = Field(
+        default=None,
+        description="Default language for transcription (e.g., 'en', 'vi').",
+    )