File size: 2,539 Bytes
b33c328 8a0efc4 b33c328 |
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 |
from random import sample
import gradio as gr
import torchaudio
import torch
import torch.nn as nn
import lightning_module
import pdb
import jiwer
# ASR part
from transformers import pipeline
p = pipeline("automatic-speech-recognition")
# WER part
transformation = jiwer.Compose([
jiwer.ToLowerCase(),
jiwer.RemoveWhiteSpace(replace_by_space=True),
jiwer.RemoveMultipleSpaces(),
jiwer.ReduceToListOfListOfWords(word_delimiter=" ")
])
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. - indices.fmod(1.)).unsqueeze(0) + round_up * indices.fmod(1.).unsqueeze(0)
return output
model = lightning_module.BaselineLightningModule.load_from_checkpoint("./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)
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
return predic_mos, trans, wer
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)
"""
iface = gr.Interface(
fn=calc_mos,
inputs=[gr.Audio(type='filepath'), gr.Textbox(placeholder="Insert referance here", label="Referance")],
outputs=[gr.Textbox("Predicted MOS"), gr.Textbox("Hypothesis"), gr.Textbox("WER")],
title="UTMOS Demo",
description=description,
allow_flagging="auto",
)
iface.launch() |