| from __future__ import annotations | |
| from typing import Any, Dict, Optional, Union | |
| from transformers import Qwen2Config, WhisperConfig | |
| from .qwen3_asr_audio_config import Qwen3ASRAudioEncoderConfig | |
| class ArkasrConfig(Qwen2Config): | |
| model_type = "arkasr" | |
| is_composition = True | |
| def __init__( | |
| self, | |
| whisper_config: Optional[Union[Dict[str, Any], WhisperConfig]] = None, | |
| qwen3_asr_audio_config: Optional[Union[Dict[str, Any], Qwen3ASRAudioEncoderConfig]] = None, | |
| adapter_type: str = "qwen3_asr_mlp_tower", | |
| merge_factor: int = 4, | |
| spec_aug: bool = False, | |
| use_rope: bool = False, | |
| max_whisper_length: int = 1500, | |
| mlp_adapter_act: str = "gelu", | |
| qwen3_asr_mlp_tower_layers: int = 4, | |
| qwen3_asr_mlp_tower_hidden_size: int = 0, | |
| qwen3_asr_mlp_tower_dropout: float = 0.0, | |
| **kwargs, | |
| ): | |
| super().__init__(**kwargs) | |
| if isinstance(whisper_config, dict): | |
| self.whisper_config = WhisperConfig(**whisper_config) | |
| elif isinstance(whisper_config, WhisperConfig): | |
| self.whisper_config = whisper_config | |
| else: | |
| self.whisper_config = WhisperConfig() | |
| if isinstance(qwen3_asr_audio_config, dict): | |
| self.qwen3_asr_audio_config = Qwen3ASRAudioEncoderConfig(**qwen3_asr_audio_config) | |
| elif isinstance(qwen3_asr_audio_config, Qwen3ASRAudioEncoderConfig): | |
| self.qwen3_asr_audio_config = qwen3_asr_audio_config | |
| else: | |
| self.qwen3_asr_audio_config = Qwen3ASRAudioEncoderConfig( | |
| num_mel_bins=128, | |
| encoder_layers=18, | |
| encoder_attention_heads=14, | |
| encoder_ffn_dim=3584, | |
| d_model=896, | |
| max_source_positions=1500, | |
| n_window=50, | |
| output_dim=1024, | |
| n_window_infer=800, | |
| conv_chunksize=500, | |
| downsample_hidden_size=480, | |
| ) | |
| self.adapter_type = adapter_type | |
| self.merge_factor = int(merge_factor) | |
| self.spec_aug = bool(spec_aug) | |
| self.use_rope = bool(use_rope) | |
| self.max_whisper_length = int(max_whisper_length) | |
| self.mlp_adapter_act = mlp_adapter_act | |
| self.qwen3_asr_mlp_tower_layers = int(qwen3_asr_mlp_tower_layers) | |
| self.qwen3_asr_mlp_tower_hidden_size = int(qwen3_asr_mlp_tower_hidden_size) | |
| self.qwen3_asr_mlp_tower_dropout = float(qwen3_asr_mlp_tower_dropout) | |
| def to_dict(self): | |
| output = super().to_dict() | |
| output["whisper_config"] = self.whisper_config.to_dict() | |
| output["qwen3_asr_audio_config"] = self.qwen3_asr_audio_config.to_dict() | |
| output["qwen3_asr_mlp_tower_layers"] = self.qwen3_asr_mlp_tower_layers | |
| output["qwen3_asr_mlp_tower_hidden_size"] = self.qwen3_asr_mlp_tower_hidden_size | |
| output["qwen3_asr_mlp_tower_dropout"] = self.qwen3_asr_mlp_tower_dropout | |
| return output | |
| __all__ = ["ArkasrConfig"] | |