The NeMo Audio Codec is a neural audio codec which compresses audio into a quantized representation. The model can be used as a vocoder for speech synthesis.
The model works with full-bandwidth 22.05kHz speech. It might have lower performance with low-bandwidth speech (e.g. 16kHz speech upsampled to 22.05kHz) or with non-speech audio.
Sample Rate
Frame Rate
Bit Rate
# Codebooks
Codebook Size
Embed Dim
FSQ Levels
22050
86.1
6.9kpbs
8
1000
32
[8, 5, 5, 5]
Model Architecture
The NeMo Audio Codec model uses symmetric convolutional encoder-decoder architecture based on HiFi-GAN. We use Finite Scalar Quantization (FSQ), with 8 codebooks and 1000 entries per codebook.
Other Properties Related to Input: 22050 Hz Mono-channel Audio
Output
Output Type: Audio
Output Format: .wav files
Output Parameters: One Dimensional (1D)
Other Properties Related to Output: 22050 Hz Mono-channel Audio
How to Use this Model
The model is available for use in the NVIDIA NeMo, and can be used as a pre-trained checkpoint for inference or for fine-tuning on another dataset.
Inference
For inference, you can follow our Audio Codec Inference Tutorial which automatically downloads the model checkpoint. Note that you will need to set the model_name parameter to "nvidia/audio-codec-22khz".
Alternatively, you can use the code below, which also handles the automatic checkpoint download:
import librosa
import torch
import soundfile as sf
from nemo.collections.tts.models import AudioCodecModel
model_name = "nvidia/audio-codec-22khz"
path_to_input_audio = ??? # path of the input audio
path_to_output_audio = ??? # path of the reconstructed output audio
nemo_codec_model = AudioCodecModel.from_pretrained(model_name).eval()
# get discrete tokens from audio
audio, _ = librosa.load(path_to_input_audio, sr=nemo_codec_model.sample_rate)
device = 'cuda' if torch.cuda.is_available() else 'cpu'
audio_tensor = torch.from_numpy(audio).unsqueeze(dim=0).to(device)
audio_len = torch.tensor([audio_tensor[0].shape[0]]).to(device)
with torch.no_grad():
encoded_tokens, encoded_len = nemo_codec_model.encode(audio=audio_tensor, audio_len=audio_len)
# Reconstruct audio from tokens
reconstructed_audio, _ = nemo_codec_model.decode(tokens=encoded_tokens, tokens_len=encoded_len)
# save reconstructed audio
output_audio = reconstructed_audio.cpu().numpy().squeeze()
sf.write(path_to_output_audio, output_audio, nemo_codec_model.sample_rate)
Training
For fine-tuning on another dataset please follow the steps available at our Audio Codec Training Tutorial. Note that you will need to set the CONFIG_FILENAME parameter to the "audio_codec_22050.yaml" config. You also will need to set pretrained_model_name to "nvidia/audio-codec-22khz".
Training, Testing, and Evaluation Datasets:
Training Datasets
The NeMo Audio Codec is trained on a total of 28.7k hrs of speech data from 105 languages.
Common Voice - 2 hours, 1356 speakers, 59 languages
Performance
We evaluate our codec using several objective audio quality metrics. We evaluate ViSQOL and PESQ for perception quality, ESTOI for intelligbility, mel spectrogram and STFT distances for spectral reconstruction accuracy, and SI-SDR for phase reconstruction accuracy. Metrics are reported on the test set for both the MLS English and CommonVoice data. The model has not been trained or evaluated on non-speech audio.
This model is for research and development only (non-commercial use) and the license to use this model is covered by the NSCLv1.
Ethical Considerations:
NVIDIA believes Trustworthy AI is a shared responsibility and we have established policies and practices to enable development for a wide array of AI applications. When downloaded or used in accordance with our terms of service, developers should work with their internal model team to ensure this model meets requirements for the relevant industry and use case and addresses unforeseen product misuse.
Please report security vulnerabilities or NVIDIA AI Concerns here.
