Datasets:

asahi417
/

experiment-audio-tokenizer

"""Basic audio reconstruction experiment."""
import os
from os.path import join as p_join
import subprocess

from datasets import load_dataset
import torch
from audiocraft.data.audio import audio_read, audio_write
from multibanddiffusion import MultiBandDiffusion

# configure experiment
cache_dir = "audio"
os.makedirs(cache_dir, exist_ok=True)


def test_files(mbd_model, num_codebooks: int = 8, skip_enhancer: bool = False):
    """Test with audio files."""
    # fetch audio
    output_audio_dir = p_join(cache_dir, "sample_audio", "original")
    os.makedirs(output_audio_dir, exist_ok=True)
    sample_audio_urls = {
        "common_voice_8_0": "https://huggingface.co/datasets/japanese-asr/ja_asr.common_voice_8_0/resolve/main/sample.flac",
        "jsut_basic5000": "https://huggingface.co/datasets/japanese-asr/ja_asr.jsut_basic5000/resolve/main/sample.flac",
        "reazonspeech_test": "https://huggingface.co/datasets/japanese-asr/ja_asr.reazonspeech_test/resolve/main/sample.flac"
    }
    for file, url in sample_audio_urls.items():
        subprocess.run(["wget", url, "-O", p_join(output_audio_dir, f"{file}.sample.flac")])
    # reconstruct audio
    output_reconstructed_dir = p_join(cache_dir, "sample_audio", f"reconstructed_{num_codebooks}codes")
    os.makedirs(output_reconstructed_dir, exist_ok=True)
    for file in sample_audio_urls.keys():
        # read audio from file
        single_file = p_join(output_audio_dir, f"{file}.sample.flac")
        wav, sr = audio_read(single_file)
        wav = wav.unsqueeze(0)
        # tokenize audio
        tokens = mbd_model.wav_to_tokens(wav, sr)
        # de-tokenize token
        re_wav, sr = mbd_model.tokens_to_wav(tokens, skip_enhancer=skip_enhancer)
        # save the reconstructed wav
        if skip_enhancer:
            output = p_join(output_reconstructed_dir, f"{file}.sample")
        else:
            output = p_join(f"{output_reconstructed_dir}.enhancer", f"{file}.sample")
        audio_write(output, re_wav[0], sr, strategy="loudness", loudness_compressor=True)


def test_hf(mbd_model, hf_dataset: str, num_codebooks: int = 8, sample_size: int = 128, batch_size: int = 32, skip_enhancer: bool = False):
    """Test with huggingface audio dataset."""
    output_dir = p_join(cache_dir, os.path.basename(hf_dataset))
    os.makedirs(output_dir, exist_ok=True)
    dataset = load_dataset(hf_dataset, split="test")
    dataset = dataset.select(range(sample_size))
    dataset = dataset.map(
        lambda batch: {k: [v] for k, v in batch.items()},
        batched=True,
        batch_size=batch_size
    )
    for data in dataset:
        # get sampling rate (all wav must be the same sampling rate)
        sr_list = [d["sampling_rate"] for d in data["audio"]]
        assert len(set(sr_list)) == 1, sr_list
        sr = sr_list[0]
        # get wav array (batch, channel, time)
        array = [d["array"] for d in data["audio"]]
        max_length = max([len(a) for a in array])
        array = [a + [0] * (max_length - len(a)) for a in array]
        wav = torch.as_tensor(array, dtype=torch.float32).unsqueeze_(1)
        # save the original wav
        for idx, one_wav in enumerate(wav):
            output = p_join(output_dir, "original", str(idx))
            audio_write(output, one_wav, sr, strategy="loudness", loudness_compressor=True)
        # tokenize audio
        tokens = mbd_model.wav_to_tokens(wav, sr)
        # de-tokenize token
        re_wav, sr = mbd_model.tokens_to_wav(tokens, skip_enhancer=skip_enhancer)
        # save the reconstructed wav
        for idx, one_wav in enumerate(re_wav):
            if skip_enhancer:
                output = p_join(output_dir, f"reconstructed_{num_codebooks}codes", str(idx))
            else:
                output = p_join(output_dir, f"reconstructed_{num_codebooks}codes.enhancer", str(idx))
            audio_write(output, one_wav, sr, strategy="loudness", loudness_compressor=True)


if __name__ == '__main__':
    # without enhancer
    for n_code in [2, 3, 4, 5, 6]:
        model = MultiBandDiffusion.from_pretrained(num_codebooks_decoder=n_code, num_codebooks_encoder=n_code)
        test_files(model, n_code, skip_enhancer=True)
        test_hf(model, "japanese-asr/ja_asr.reazonspeech_test", num_codebooks=n_code, sample_size=64, batch_size=16, skip_enhancer=True)
        test_hf(model, "japanese-asr/ja_asr.jsut_basic5000", num_codebooks=n_code, sample_size=64, batch_size=16, skip_enhancer=True)
    # with enhancer
    n_code = 3
    model = MultiBandDiffusion.from_pretrained(num_codebooks_decoder=n_code, num_codebooks_encoder=n_code)
    test_files(model, n_code)
    test_hf(model, "japanese-asr/ja_asr.reazonspeech_test", num_codebooks=n_code, sample_size=64, batch_size=16)
    test_hf(model, "japanese-asr/ja_asr.jsut_basic5000", num_codebooks=n_code, sample_size=64, batch_size=16)

	"""Basic audio reconstruction experiment."""
	import os
	from os.path import join as p_join
	import subprocess

	from datasets import load_dataset
	import torch
	from audiocraft.data.audio import audio_read, audio_write
	from multibanddiffusion import MultiBandDiffusion

	# configure experiment
	cache_dir = "audio"
	os.makedirs(cache_dir, exist_ok=True)


	def test_files(mbd_model, num_codebooks: int = 8, skip_enhancer: bool = False):
	"""Test with audio files."""
	# fetch audio
	output_audio_dir = p_join(cache_dir, "sample_audio", "original")
	os.makedirs(output_audio_dir, exist_ok=True)
	sample_audio_urls = {
	"common_voice_8_0": "https://huggingface.co/datasets/japanese-asr/ja_asr.common_voice_8_0/resolve/main/sample.flac",
	"jsut_basic5000": "https://huggingface.co/datasets/japanese-asr/ja_asr.jsut_basic5000/resolve/main/sample.flac",
	"reazonspeech_test": "https://huggingface.co/datasets/japanese-asr/ja_asr.reazonspeech_test/resolve/main/sample.flac"
	}
	for file, url in sample_audio_urls.items():
	subprocess.run(["wget", url, "-O", p_join(output_audio_dir, f"{file}.sample.flac")])
	# reconstruct audio
	output_reconstructed_dir = p_join(cache_dir, "sample_audio", f"reconstructed_{num_codebooks}codes")
	os.makedirs(output_reconstructed_dir, exist_ok=True)
	for file in sample_audio_urls.keys():
	# read audio from file
	single_file = p_join(output_audio_dir, f"{file}.sample.flac")
	wav, sr = audio_read(single_file)
	wav = wav.unsqueeze(0)
	# tokenize audio
	tokens = mbd_model.wav_to_tokens(wav, sr)
	# de-tokenize token
	re_wav, sr = mbd_model.tokens_to_wav(tokens, skip_enhancer=skip_enhancer)
	# save the reconstructed wav
	if skip_enhancer:
	output = p_join(output_reconstructed_dir, f"{file}.sample")
	else:
	output = p_join(f"{output_reconstructed_dir}.enhancer", f"{file}.sample")
	audio_write(output, re_wav[0], sr, strategy="loudness", loudness_compressor=True)


	def test_hf(mbd_model, hf_dataset: str, num_codebooks: int = 8, sample_size: int = 128, batch_size: int = 32, skip_enhancer: bool = False):
	"""Test with huggingface audio dataset."""
	output_dir = p_join(cache_dir, os.path.basename(hf_dataset))
	os.makedirs(output_dir, exist_ok=True)
	dataset = load_dataset(hf_dataset, split="test")
	dataset = dataset.select(range(sample_size))
	dataset = dataset.map(
	lambda batch: {k: [v] for k, v in batch.items()},
	batched=True,
	batch_size=batch_size
	)
	for data in dataset:
	# get sampling rate (all wav must be the same sampling rate)
	sr_list = [d["sampling_rate"] for d in data["audio"]]
	assert len(set(sr_list)) == 1, sr_list
	sr = sr_list[0]
	# get wav array (batch, channel, time)
	array = [d["array"] for d in data["audio"]]
	max_length = max([len(a) for a in array])
	array = [a + [0] * (max_length - len(a)) for a in array]
	wav = torch.as_tensor(array, dtype=torch.float32).unsqueeze_(1)
	# save the original wav
	for idx, one_wav in enumerate(wav):
	output = p_join(output_dir, "original", str(idx))
	audio_write(output, one_wav, sr, strategy="loudness", loudness_compressor=True)
	# tokenize audio
	tokens = mbd_model.wav_to_tokens(wav, sr)
	# de-tokenize token
	re_wav, sr = mbd_model.tokens_to_wav(tokens, skip_enhancer=skip_enhancer)
	# save the reconstructed wav
	for idx, one_wav in enumerate(re_wav):
	if skip_enhancer:
	output = p_join(output_dir, f"reconstructed_{num_codebooks}codes", str(idx))
	else:
	output = p_join(output_dir, f"reconstructed_{num_codebooks}codes.enhancer", str(idx))
	audio_write(output, one_wav, sr, strategy="loudness", loudness_compressor=True)


	if __name__ == '__main__':
	# without enhancer
	for n_code in [2, 3, 4, 5, 6]:
	model = MultiBandDiffusion.from_pretrained(num_codebooks_decoder=n_code, num_codebooks_encoder=n_code)
	test_files(model, n_code, skip_enhancer=True)
	test_hf(model, "japanese-asr/ja_asr.reazonspeech_test", num_codebooks=n_code, sample_size=64, batch_size=16, skip_enhancer=True)
	test_hf(model, "japanese-asr/ja_asr.jsut_basic5000", num_codebooks=n_code, sample_size=64, batch_size=16, skip_enhancer=True)
	# with enhancer
	n_code = 3
	model = MultiBandDiffusion.from_pretrained(num_codebooks_decoder=n_code, num_codebooks_encoder=n_code)
	test_files(model, n_code)
	test_hf(model, "japanese-asr/ja_asr.reazonspeech_test", num_codebooks=n_code, sample_size=64, batch_size=16)
	test_hf(model, "japanese-asr/ja_asr.jsut_basic5000", num_codebooks=n_code, sample_size=64, batch_size=16)