src/bigru_t/multimodal/audio_encoder.py

"""Xavante - audio_encoder.py - Encoder de audio (estilo Whisper leve)."""
from __future__ import annotations

import logging

import torch
import torch.nn as nn

logger = logging.getLogger(__name__)


class AudioEncoder(nn.Module):
    """Convolucoes 1D + Linear para audio bruto ou mel-spec."""

    def __init__(self, d_model: int = 512, in_channels: int = 1, n_mels: int = 80):
        super().__init__()
        in_ch = in_channels if in_channels > 1 else n_mels
        self.conv = nn.Sequential(
            nn.Conv1d(in_ch, 64, 5, 2, 2),
            nn.GELU(),
            nn.Conv1d(64, 128, 5, 2, 2),
            nn.GELU(),
            nn.Conv1d(128, 256, 5, 2, 2),
            nn.GELU(),
            nn.AdaptiveAvgPool1d(1),
        )
        self.proj = nn.Linear(256, d_model)

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        # x: [B, C, T] ou [B, n_mels, T]
        feat = self.conv(x).squeeze(-1)
        return self.proj(feat)


__all__ = ["AudioEncoder"]