File size: 2,629 Bytes
20857b0
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
import torch

from .base import AudioTokenizer


class EnCodecAudioCodec(AudioTokenizer):
    def __init__(self, device: str = "cpu", bandwidth: float = 6.0):
        self.device = torch.device(device)
        self._bandwidth = bandwidth
        self._model = None
        self._loaded = False

    def _lazy_load(self):
        if self._loaded:
            return
        try:
            from encodec import EncodecModel
            self._model = EncodecModel.encodec_model_24khz()
            self._model.set_target_bandwidth(self._bandwidth)
            self._model.to(self.device)
            self._model.eval()
            self._loaded = True
        except ImportError:
            raise ImportError("encodec package not installed; run: pip install encodec")
        except Exception as e:
            raise RuntimeError(f"failed to load EnCodec: {e}")

    @torch.inference_mode()
    def encode(self, audio: torch.Tensor) -> torch.Tensor:
        self._lazy_load()
        audio = audio.to(self.device)
        if audio.dim() == 1:
            audio = audio.unsqueeze(0).unsqueeze(0)
        elif audio.dim() == 2:
            audio = audio.unsqueeze(1)
        from encodec.utils import convert_audio
        audio = convert_audio(audio, self._model.sample_rate, self._model.sample_rate, self._model.channels)
        frames = self._model.encode(audio)
        codes = torch.cat([f[0] for f in frames], dim=-1)
        return codes

    @torch.inference_mode()
    def decode(self, tokens: torch.Tensor) -> torch.Tensor:
        self._lazy_load()
        if tokens.dim() == 2:
            tokens = tokens.unsqueeze(0)
        frames = self._model.decode([(tokens, None)])
        return frames[0]

    @property
    def sample_rate(self) -> int:
        return 24000

    @property
    def num_codebooks(self) -> int:
        return 8

    @property
    def name(self) -> str:
        return f"encodec-{self._bandwidth}kbps"


class DummyAudioCodec(AudioTokenizer):
    def __init__(self, device: str = "cpu"):
        self.device = torch.device(device)

    def encode(self, audio: torch.Tensor) -> torch.Tensor:
        B = audio.shape[0] if audio.dim() > 1 else 1
        return torch.randint(0, 2048, (4, B * 10), dtype=torch.int32, device=self.device)

    def decode(self, tokens: torch.Tensor) -> torch.Tensor:
        length = tokens.shape[-1] * 320
        return torch.randn(1, length, device=self.device)

    @property
    def sample_rate(self) -> int:
        return 32000

    @property
    def num_codebooks(self) -> int:
        return 4

    @property
    def name(self) -> str:
        return "dummy"