SoundStream: An End-to-End Neural Audio Codec

TL;DR

SoundStream presents an end-to-end neural audio codec that uses a fully convolutional encoder/decoder coupled with a residual vector quantizer, trained with reconstruction and adversarial losses. A novel quantizer dropout enables a single model to operate across 3–18 kbps without architectural changes, delivering high perceptual quality across speech, music, and general audio, and running in real time on a smartphone CPU. The approach is extended to joint compression and enhancement via FiLM conditioning, allowing denoising without added latency and achieving bitrate savings through encoder-side denoising. Empirical results show SoundStream outperforms Opus at low bitrates and approaches EVS, with strong bitrate scalability and efficient real-time performance on resource-limited hardware.

Abstract

We present SoundStream, a novel neural audio codec that can efficiently compress speech, music and general audio at bitrates normally targeted by speech-tailored codecs. SoundStream relies on a model architecture composed by a fully convolutional encoder/decoder network and a residual vector quantizer, which are trained jointly end-to-end. Training leverages recent advances in text-to-speech and speech enhancement, which combine adversarial and reconstruction losses to allow the generation of high-quality audio content from quantized embeddings. By training with structured dropout applied to quantizer layers, a single model can operate across variable bitrates from 3kbps to 18kbps, with a negligible quality loss when compared with models trained at fixed bitrates. In addition, the model is amenable to a low latency implementation, which supports streamable inference and runs in real time on a smartphone CPU. In subjective evaluations using audio at 24kHz sampling rate, SoundStream at 3kbps outperforms Opus at 12kbps and approaches EVS at 9.6kbps. Moreover, we are able to perform joint compression and enhancement either at the encoder or at the decoder side with no additional latency, which we demonstrate through background noise suppression for speech.

Figures (8)

• Figure 1: SoundStream@$3\,$kbps vs. state-of-the-art codecs.
• Figure 2: SoundStream model architecture. A convolutional encoder produces a latent representation of the input audio samples, which is quantized using a variable number $n_q$ of residual vector quantizers (RVQ). During training, the model parameters are optimized using a combination of reconstruction and adversarial losses. An optional conditioning input can be used to indicate whether background noise has to be removed from the audio. When deploying the model, the encoder and quantizer on a transmitter client send the compressed bitstream to a receiver client that can then decode the audio signal.
• Figure 3: Encoder and decoder model architecture.
• Figure 4: STFT-based discriminator architecture.
• Figure 5: Subjective evaluation results. Error bars denote $95\%$ confidence intervals.