File size: 8,485 Bytes
a1fe393
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
# Get Transcription, WER and PPM 
"""
TODO:
    [DONE]: Automatic generating Config
"""

import yaml
import argparse
import sys
from pathlib import Path

sys.path.append("./src")
import lightning_module
from UV import plot_UV, get_speech_interval
from transformers import pipeline
from rich.progress import track
from rich import print as rprint
import numpy as np
import jiwer
import pdb
import torch.nn as nn
import torch
import torchaudio
import gradio as gr
from sys import flags
from random import sample
from transformers import Wav2Vec2Processor, Wav2Vec2ForCTC

# root_path = Path(__file__).parents[1]

class ChangeSampleRate(nn.Module):
    def __init__(self, input_rate: int, output_rate: int):
        super().__init__()
        self.output_rate = output_rate
        self.input_rate = input_rate

    def forward(self, wav: torch.tensor) -> torch.tensor:
        # Only accepts 1-channel waveform input
        wav = wav.view(wav.size(0), -1)
        new_length = wav.size(-1) * self.output_rate // self.input_rate
        indices = torch.arange(new_length) * (
            self.input_rate / self.output_rate
        )
        round_down = wav[:, indices.long()]
        round_up = wav[:, (indices.long() + 1).clamp(max=wav.size(-1) - 1)]
        output = round_down * (1.0 - indices.fmod(1.0)).unsqueeze(
            0
        ) + round_up * indices.fmod(1.0).unsqueeze(0)
        return output


model = lightning_module.BaselineLightningModule.load_from_checkpoint(
    "./src/epoch=3-step=7459.ckpt"
).eval()


def calc_mos(audio_path, ref):
    wav, sr = torchaudio.load(audio_path)
    osr = 16_000
    batch = wav.unsqueeze(0).repeat(10, 1, 1)
    csr = ChangeSampleRate(sr, osr)
    out_wavs = csr(wav)
    # ASR
    trans = p(audio_path)["text"]
    # WER
    wer = jiwer.wer(
        ref,
        trans,
        truth_transform=transformation,
        hypothesis_transform=transformation,
    )
    # MOS
    batch = {
        "wav": out_wavs,
        "domains": torch.tensor([0]),
        "judge_id": torch.tensor([288]),
    }
    with torch.no_grad():
        output = model(batch)
    predic_mos = output.mean(dim=1).squeeze().detach().numpy() * 2 + 3
    # Phonemes per minute (PPM)
    with torch.no_grad():
        logits = phoneme_model(out_wavs).logits
    phone_predicted_ids = torch.argmax(logits, dim=-1)
    phone_transcription = processor.batch_decode(phone_predicted_ids)
    lst_phonemes = phone_transcription[0].split(" ")
    wav_vad = torchaudio.functional.vad(wav, sample_rate=sr)
    ppm = len(lst_phonemes) / (wav_vad.shape[-1] / sr) * 60
    # if float(predic_mos) >= 3.0:
    #     torchaudio.save("good.wav", wav,sr)

    return predic_mos, trans, wer, phone_transcription, ppm

if __name__ == "__main__":
    # Argparse
    parser = argparse.ArgumentParser(
        prog="get_ref_PPM",
        description="Generate Phoneme per Minute (and Voice/Unvoice plot)",
        epilog="",
    )
    parser.add_argument(
        "--tag",
        type=str,
        default=None,
        required=False,
        help="ID tag for output *.csv",
    )

    parser.add_argument("--ref_txt", type=str, required=True, help="Reference TXT")
    parser.add_argument(
        "--ref_wavs", type=str, required=True, help="Reference WAVs"
    )

    parser.add_argument(
        "--output_dir",
        type=str,
        required=True,
        help="Output Directory for *.csv",
    )
    parser.add_argument(
        "--to_config",
        choices=["True", "False"],
        default="False",
        help="Generating Config from .txt and wavs/*wav",
    )

    parser.add_argument(
        "--UV_flag",
        choices=["True", "False"],
        default="False",
        help="Toggle for U/V plot",
    )
    parser.add_argument(
        "--UV_thre", type=float, default=40, help="U/V threshold dB"
    )
    args = parser.parse_args()

    refs = np.loadtxt(args.ref_txt, delimiter="\n", dtype="str")
    refs_ids = [x.split()[0] for x in refs]
    refs_txt = [" ".join(x.split()[1:]) for x in refs]
    ref_wavs = [str(x) for x in sorted(Path(args.ref_wavs).glob("**/*.wav"))]
    # pdb.set_trace()
    try:
        len(refs) == len(ref_wavs)
    except ValueError:
        print("Error: Text and Wavs don't match")
        exit()

    # ASR part
    p = pipeline("automatic-speech-recognition")

    # WER part
    transformation = jiwer.Compose(
        [
            jiwer.ToLowerCase(),
            jiwer.RemoveWhiteSpace(replace_by_space=True),
            jiwer.RemoveMultipleSpaces(),
            jiwer.ReduceToListOfListOfWords(word_delimiter=" "),
        ]
    )

    # WPM part
    processor = Wav2Vec2Processor.from_pretrained(
        "facebook/wav2vec2-xlsr-53-espeak-cv-ft"
    )
    phoneme_model = Wav2Vec2ForCTC.from_pretrained(
        "facebook/wav2vec2-xlsr-53-espeak-cv-ft"
    )
    # phoneme_model =  pipeline(model="facebook/wav2vec2-xlsr-53-espeak-cv-ft")

    description = """
    MOS prediction demo using UTMOS-strong w/o phoneme encoder model, \
        which is trained on the main track dataset.
        This demo only accepts .wav format. Best at 16 kHz sampling rate.

    Paper is available [here](https://arxiv.org/abs/2204.02152)

    Add ASR based on wav2vec-960, currently only English available.
    Add WER interface.
    """

    referance_id = gr.Textbox(
        value="ID", placeholder="Utter ID", label="Reference_ID"
    )
    referance_textbox = gr.Textbox(
        value="", placeholder="Input reference here", label="Reference"
    )
    # Set up interface
    result = []
    result.append("id, pred_mos, trans, wer, pred_phone, ppm")

    if args.UV_flag == "False":
        for id, x, y in track(
            zip(refs_ids, ref_wavs, refs_txt),
            total=len(refs_ids),
            description="Loading references information",
        ):
            predic_mos, trans, wer, phone_transcription, ppm = calc_mos(x, y)
            record = ",".join(
                [
                    id,
                    str(predic_mos),
                    str(trans),
                    str(wer),
                    str(phone_transcription),
                    str(ppm),
                ]
            )
            result.append(record)

    elif args.UV_flag == "True":
        fig_tardir = Path(args.ref_wavs) / Path("PPM_figs")
        Path.mkdir(Path(args.ref_wavs) / Path("PPM_figs"), exist_ok=True)

        for id, x, y in track(
            zip(refs_ids, ref_wavs, refs_txt),
            total=len(refs_ids),
            description="Loading references information",
        ):
            # UV ploting
            wav, sr = torchaudio.load(x)
            wav_vad = torchaudio.functional.vad(wav, sample_rate=sr)
            a_h, p_h = get_speech_interval(wav_vad.numpy(), db=args.UV_thre)
            fig_h = plot_UV(wav_vad.numpy().squeeze(), a_h, sr=sr)
            fig_h.savefig(Path(fig_tardir) / Path(id + ".png"), dpi=200)
            # Acoustic calculation
            predic_mos, trans, wer, phone_transcription, ppm = calc_mos(x, y)
            record = ",".join(
                [
                    id,
                    str(predic_mos),
                    str(trans),
                    str(wer),
                    str(phone_transcription),
                    str(ppm),
                ]
            )
            result.append(record)
    # Output
    if args.tag == None:
        args.tag = Path(args.ref_wavs).stem
    # Make output_dir
    # pdb.set_trace()
    Path.mkdir(Path(args.output_dir), exist_ok=True)
    # pdb.set_trace()
    with open("%s/%s.csv" % (args.output_dir, args.tag), "w") as f:
        print("\n".join(result), file=f)

    # Generating config
    if args.to_config == "True":
        config_dict = {
            "exp_id": args.tag,
            "ref_txt": args.ref_txt,
            "ref_feature": "%s/%s.csv" % (args.output_dir, args.tag),
            "ref_wavs": args.ref_wavs,
            "thre": {
                "minppm": 100,
                "maxppm": 100,
                "WER": 0.1,
                "AUTOMOS": 4.0,
            },
            "auth": {"username": None, "password": None},
        }
        with open("./config/%s.yaml" % args.tag, "w") as config_f:
            rprint("Dumping as config ./config/%s.yaml" % args.tag)
            rprint(config_dict)
            yaml.dump(config_dict, stream=config_f)
            rprint("Change parameter ./config/%s.yaml if necessary" % args.tag)
    print("Reference Dumping Finished")