Step-Audio2/token2wav.py at main · stepfun-ai/Step-Audio2 · GitHub

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
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
import io

import torch
import torchaudio
import s3tokenizer
import onnxruntime
import numpy as np

import torchaudio.compliance.kaldi as kaldi
from flashcosyvoice.modules.hifigan import HiFTGenerator
from flashcosyvoice.utils.audio import mel_spectrogram
from hyperpyyaml import load_hyperpyyaml

def fade_in_out(fade_in_mel:torch.Tensor, fade_out_mel:torch.Tensor, window:torch.Tensor):
    """perform fade_in_out in tensor style
    """
    mel_overlap_len = int(window.shape[0] / 2)
    fade_in_mel = fade_in_mel.clone()
    fade_in_mel[..., :mel_overlap_len] = \
        fade_in_mel[..., :mel_overlap_len] * window[:mel_overlap_len] + \
        fade_out_mel[..., -mel_overlap_len:] * window[mel_overlap_len:]
    return fade_in_mel


class Token2wav():

    def __init__(self, model_path, float16=False):
        self.float16 = float16

        self.audio_tokenizer = s3tokenizer.load_model(f"{model_path}/speech_tokenizer_v2_25hz.onnx").cuda().eval()

        option = onnxruntime.SessionOptions()
        option.graph_optimization_level = onnxruntime.GraphOptimizationLevel.ORT_ENABLE_ALL
        option.intra_op_num_threads = 1
        self.spk_model = onnxruntime.InferenceSession(f"{model_path}/campplus.onnx", sess_options=option, providers=["CPUExecutionProvider"])

        with open(f"{model_path}/flow.yaml", "r") as f:
            configs = load_hyperpyyaml(f)
            self.flow = configs['flow']
        if float16:
            self.flow.half()
        self.flow.load_state_dict(torch.load(f"{model_path}/flow.pt", map_location="cpu", weights_only=True), strict=True)
        self.flow.cuda().eval()

        self.hift = HiFTGenerator()
        hift_state_dict = {k.replace('generator.', ''): v for k, v in torch.load(f"{model_path}/hift.pt", map_location="cpu", weights_only=True).items()}
        self.hift.load_state_dict(hift_state_dict, strict=True)
        self.hift.cuda().eval()

        self.cache = {}

        # stream conf
        self.mel_cache_len = 8  # hard-coded, 160ms
        self.source_cache_len = int(self.mel_cache_len * 480)   # 50hz mel -> 24kHz wave
        self.speech_window = torch.from_numpy(np.hamming(2 * self.source_cache_len)).cuda()

        # hifigan cache
        self.hift_cache_dict = {}


    def _prepare_prompt(self, prompt_wav):
        audio = s3tokenizer.load_audio(prompt_wav, sr=16000)  # [T]
        mels = s3tokenizer.log_mel_spectrogram(audio)
        mels, mels_lens = s3tokenizer.padding([mels])
        prompt_speech_tokens, prompt_speech_tokens_lens = self.audio_tokenizer.quantize(mels.cuda(), mels_lens.cuda())

        spk_feat = kaldi.fbank(audio.unsqueeze(0), num_mel_bins=80, dither=0, sample_frequency=16000)
        spk_feat = spk_feat - spk_feat.mean(dim=0, keepdim=True)
        spk_emb = torch.tensor(self.spk_model.run(
            None, {self.spk_model.get_inputs()[0].name: spk_feat.unsqueeze(dim=0).cpu().numpy()}
        )[0], device='cuda')

        audio, sample_rate = torchaudio.load(prompt_wav, backend='soundfile')
        audio = audio.mean(dim=0, keepdim=True)  # [1, T]
        if sample_rate != 24000:
            audio = torchaudio.transforms.Resample(orig_freq=sample_rate, new_freq=24000)(audio)
        prompt_mel = mel_spectrogram(audio).transpose(1, 2).squeeze(0)  # [T, num_mels]
        prompt_mels = prompt_mel.unsqueeze(0).cuda()
        prompt_mels_lens = torch.tensor([prompt_mels.shape[1]], dtype=torch.int32, device='cuda')
        prompt_mels = torch.nn.functional.pad(prompt_mels, (0, 0, 0, prompt_speech_tokens.shape[1] * self.flow.up_rate - prompt_mels.shape[1]), mode='replicate')
        return prompt_speech_tokens, prompt_speech_tokens_lens, spk_emb, prompt_mels, prompt_mels_lens

    def __call__(self, generated_speech_tokens, prompt_wav):
        if prompt_wav not in self.cache:
            self.cache[prompt_wav] = self._prepare_prompt(prompt_wav)
        prompt_speech_tokens, prompt_speech_tokens_lens, spk_emb, prompt_mels, prompt_mels_lens = self.cache[prompt_wav]

        generated_speech_tokens = torch.tensor([generated_speech_tokens], dtype=torch.int32, device='cuda')
        generated_speech_tokens_lens = torch.tensor([generated_speech_tokens.shape[1]], dtype=torch.int32, device='cuda')

        with torch.amp.autocast("cuda", dtype=torch.float16 if self.float16 else torch.float32):
            mel = self.flow.inference(generated_speech_tokens, generated_speech_tokens_lens,
                prompt_speech_tokens, prompt_speech_tokens_lens,
                prompt_mels, prompt_mels_lens, spk_emb, 10)

        wav, _ = self.hift(speech_feat=mel)
        output = io.BytesIO()
        torchaudio.save(output, wav.cpu(), sample_rate=24000, format='wav')

        return output.getvalue()

    def set_stream_cache(self, prompt_wav):
        if prompt_wav not in self.cache:
            self.cache[prompt_wav] = self._prepare_prompt(prompt_wav)
        prompt_speech_tokens, prompt_speech_tokens_lens, spk_emb, prompt_mels, prompt_mels_lens = self.cache[prompt_wav]
        self.stream_cache = self.flow.setup_cache(
            torch.cat([prompt_speech_tokens, prompt_speech_tokens[:, :3]], dim=1),
            prompt_mels, spk_emb, n_timesteps=10)

        # hift cache
        self.hift_cache_dict = dict(
            mel = torch.zeros(1, prompt_mels.shape[2], 0, device='cuda'),
            source = torch.zeros(1, 1, 0, device='cuda'),
            speech = torch.zeros(1, 0, device='cuda'),
        )


    def stream(self, generated_speech_tokens, prompt_wav, last_chunk=False):
        if prompt_wav not in self.cache:
            self.cache[prompt_wav] = self._prepare_prompt(prompt_wav)
        prompt_speech_tokens, prompt_speech_tokens_lens, spk_emb, prompt_mels, prompt_mels_lens = self.cache[prompt_wav]

        generated_speech_tokens = torch.tensor([generated_speech_tokens], dtype=torch.int32, device='cuda')
        generated_speech_tokens_lens = torch.tensor([generated_speech_tokens.shape[1]], dtype=torch.int32, device='cuda')

        if self.stream_cache is None:
            raise ValueError("stream_cache is not set")

        with torch.amp.autocast("cuda", dtype=torch.float16 if self.float16 else torch.float32):
            chunk_mel, self.stream_cache = self.flow.inference_chunk(
                token=generated_speech_tokens,
                spk=spk_emb,
                cache=self.stream_cache,
                last_chunk=last_chunk,
                n_timesteps=10,
            )
        if self.stream_cache['estimator_att_cache'].shape[4] > (prompt_mels.shape[1] + 100):
            self.stream_cache['estimator_att_cache'] = torch.cat([
                self.stream_cache['estimator_att_cache'][:, :, :, :, :prompt_mels.shape[1]],
                self.stream_cache['estimator_att_cache'][:, :, :, :, -100:],
            ], dim=4)

        # vocoder cache
        hift_cache_mel = self.hift_cache_dict['mel']
        hift_cache_source = self.hift_cache_dict['source']
        hift_cache_speech = self.hift_cache_dict['speech']
        mel = torch.concat([hift_cache_mel, chunk_mel], dim=2)

        speech, source = self.hift(mel, hift_cache_source)

        # overlap speech smooth
        if hift_cache_speech.shape[-1] > 0:
            speech = fade_in_out(speech, hift_cache_speech, self.speech_window)

        # update vocoder cache
        self.hift_cache_dict = dict(
            mel = mel[..., -self.mel_cache_len:].clone().detach(),
            source = source[:, :, -self.source_cache_len:].clone().detach(),
            speech = speech[:, -self.source_cache_len:].clone().detach(),
        )
        if not last_chunk:
            speech = speech[:, :-self.source_cache_len]

        wav_np = speech.cpu().numpy()
        # Clip to [-1, 1] to avoid overflow, then scale to int16
        wav_np = np.clip(wav_np, -1.0, 1.0)
        wav_int16 = (wav_np * 32767.0).astype('<i2')  # 16-bit little-endian PCM
        pcm_bytes = wav_int16.tobytes()
        return pcm_bytes

if __name__ == '__main__':
    token2wav = Token2wav('Step-Audio-2-mini/token2wav')

    tokens = [1493, 4299, 4218, 2049, 528, 2752, 4850, 4569, 4575, 6372, 2127, 4068, 2312, 4993, 4769, 2300, 226, 2175, 2160, 2152, 6311, 6065, 4859, 5102, 4615, 6534, 6426, 1763, 2249, 2209, 5938, 1725, 6048, 3816, 6058, 958, 63, 4460, 5914, 2379, 735, 5319, 4593, 2328, 890, 35, 751, 1483, 1484, 1483, 2112, 303, 4753, 2301, 5507, 5588, 5261, 5744, 5501, 2341, 2001, 2252, 2344, 1860, 2031, 414, 4366, 4366, 6059, 5300, 4814, 5092, 5100, 1923, 3054, 4320, 4296, 2148, 4371, 5831, 5084, 5027, 4946, 4946, 2678, 575, 575, 521, 518, 638, 1367, 2804, 3402, 4299]
    audio = token2wav(tokens, 'assets/default_male.wav')
    with open('assets/give_me_a_brief_introduction_to_the_great_wall.wav', 'wb') as f:
        f.write(audio)