We introduce LMCodec, a causal neural speech codec that provides high quality audio at very low bitrates. The backbone of the system is a causal convolutional codec that encodes audio into a hierarchy of coarse-to-fine tokens using residual vector quantization. LMCodec trains a Transformer language model to predict the fine tokens from the coarse ones in a generative fashion, allowing for the transmission of fewer codes. A second Transformer predicts the uncertainty of the next codes given the past transmitted codes, and is used to perform conditional entropy coding. A MUSHRA subjective test was conducted and shows that the quality is comparable to reference codecs at higher bitrates. Example audio is available at https://mjenrungrot.github.io/chrome-media-audio-papers/publications/lmcodec.

We present a multi-channel audio signal generation scheme based on machine-learning and probabilistic modeling. We start from modeling a multi-channel single-source signal. Such signals are naturally modeled as a single-channel reference signal and a spatial-arrangement (SA) model specified by an SA parameter sequence. We focus on the SA model and assume that the reference signal is described by some parameter sequence. The SA model parameters are described with a learned probability distribution that is conditioned by the reference-signal parameter sequence and, optionally, an SA conditioning sequence. If present, the SA conditioning sequence specifies a signal class or a specific signal. The single-source method can be used for multi-source signals by applying source separation or by using an SA model that operates on nonoverlapping frequency bands. Our GAN-based stereo coding implementation of the latter approach shows that our paradigm facilitates plausible high-quality rendering at a low bit rate for the SA conditioning.

Speech coding facilitates the transmission of speech over lowbandwidth networks with minimal distortion.Neural-network based speech codecs have recently demonstrated significant improvements in quality over traditional approaches.While this new generation of codecs is capable of synthesizing highfidelity speech, their use of recurrent or convolutional layers often restricts their effective receptive fields, which prevents them from compressing speech efficiently.We propose to further reduce the bitrate of neural speech codecs through the use of pretrained Transformers, capable of exploiting long-range dependencies in the input signal due to their inductive bias.As such, we use a pretrained Transformer in tandem with a convolutional encoder, which is trained end-to-end with a quantizer and a generative adversarial net decoder.Our numerical experiments show that supplementing the convolutional encoder of a neural speech codec with Transformer speech embeddings yields a speech codec with a bitrate of 600 bps that outperforms the original neural speech codec in synthesized speech quality when trained at the same bitrate.Subjective human evaluations suggest that the quality of the resulting codec is comparable or better than that of conventional codecs operating at three to four times the rate.

Abstract Speech coding has been shown to achieve good speech quality using either waveform matching or parametric reconstruction. For very low bit rate streams, recently developed generative speech models can reconstruct high‐quality wideband speech from the bit streams of standard parametric encoders at less than 3 kb/s. Generative codecs produce high‐quality speech based on synthesising speech from a DNN and the parametric input. Existing objective speech quality models (e.g., ViSQOL and POLQA) cannot be used to accurately evaluate the quality of coded speech from generative models as they penalise based on signal differences not apparent in subjective listening test results. This paper presents WARP‐Q, a full‐reference objective speech quality metric that uses a dynamic time warping cost for MFCC representations of the signals. It is robust to low perceptual signal changes introduced by low bit rate neural vocoders. An evaluation using waveform matching, parametric, and generative neural vocoder‐based codecs as well as channel and environmental noise shows that WARP‐Q has better correlation and codec quality ranking for novel codecs compared to traditional metrics as well as the versatility of capturing other types of degradations, such as additive noise and transmission channel degradations.

Good speech quality has been achieved using waveform matching and parametric reconstruction coders. Recently developed very low bit rate generative codecs can reconstruct high quality wideband speech with bit streams less than 3 kb/s. These codecs use a DNN with parametric input to synthesise high quality speech outputs. Existing objective speech quality models (e.g., POLQA, ViSQOL) do not accurately predict the quality of coded speech from these generative models underestimating quality due to signal differences not highlighted in subjective listening tests. We present WARP-Q, a full-reference objective speech quality metric that uses dynamic time warping cost for MFCC speech representations. It is robust to small perceptual signal changes. Evaluation using waveform matching, parametric and generative neural vocoder based codecs as well as channel and environmental noise shows that WARP-Q has better correlation and codec quality ranking for novel codecs compared to traditional metrics in addition to versatility for general quality assessment scenarios.

The recent emergence of machine-learning based generative models for speech suggests a significant reduction in bit rate for speech codecs is possible. However, the performance of generative models deteriorates significantly with the distortions present in real-world input signals. We argue that this deterioration is due to the sensitivity of the maximum likelihood criterion to outliers and the ineffectiveness of modeling a sum of independent signals with a single autoregressive model. We introduce predictive-variance regularization to reduce the sensitivity to outliers, resulting in a significant increase in performance. We show that noise reduction to remove unwanted signals can significantly increase performance. We provide extensive subjective performance evaluations that show that our system based on generative modeling provides state-of-the-art coding performance at 3 kb/s for real-world speech signals at reasonable computational complexity.

Rapid advances in machine-learning based generative modeling of speech make its use in speech coding attractive. However, the current performance of such models drops rapidly with noise contamination of the input, preventing use in practical applications. We present a new speech-coding scheme that is based on features that are robust to the distortions occurring in speech-coder input signals. To this purpose, we encourage the feature encoder to provide the same independent features for each of a set of linguistically equivalent signals, obtained by adding various noises to a common clean signal. The independent features, subjected to scalar quantization, are used as a conditioning vector sequence for WaveNet. Our experiments show that a 1.8 kb/s implementation of the resulting coder provides state-of-the-art performance for clean signals, and is additionally robust to noisy input.

Estimation of perceptual quality in audio and speech is possible using a variety of methods. The combined v3 release of ViSQOL and ViSQOLAudio (for speech and audio, respectively,) provides improvements upon previous versions, in terms of both design and usage. As an open source C++ library or binary with permissive licensing, ViSQOL can now be deployed beyond the research context into production usage. The feedback from internal production teams at Google has helped to improve this new release, and serves to show cases where it is most applicable, as well as to highlight limitations. The new model is benchmarked against real-world data for evaluation purposes. The trends and direction of future work is discussed.

This study compares the performances of different algorithms for coding speech at low bit rates. In addition to widely deployed traditional vocoders, a selection of recently developed generative-model-based coders at different bit rates are contrasted. Performance analysis of the coded speech is evaluated for different quality aspects: accuracy of pitch periods estimation, the word error rates for automatic speech recognition, and the influence of speaker gender and coding delays. A number of performance metrics of speech samples taken from a publicly available database were compared with subjective scores. Results from subjective quality assessment do not correlate well with existing full reference speech quality metrics. The results provide valuable insights into aspects of the speech signal that will be used to develop a novel metric to accurately predict speech quality from generative-model-based coders.

In recent years, deep networks have led to dramatic improvements in speech enhancement by framing it as a data-driven pattern recognition problem. In many modern enhancement systems, large amounts of data are used to train a deep network to estimate masks for complex-valued short-time Fourier transforms (STFTs) to suppress noise and preserve speech. However, current masking approaches often neglect two important constraints: STFT consistency and mixture consistency. Without STFT consistency, the system's output is not necessarily the STFT of a time-domain signal, and without mixture consistency, the sum of the estimated sources does not necessarily equal the input mixture. Furthermore, the only previous approaches that apply mixture consistency use real-valued masks; mixture consistency has been ignored for complex-valued masks.

We define salient features as features that are shared by signals that are defined as being equivalent by a system designer. The definition allows the designer to contribute qualitative information. We aim to find salient features that are useful as conditioning for generative networks. We extract salient features by jointly training a set of clones of an encoder network. Each network clone receives as input a different signal from a set of equivalent signals. The objective function encourages the network clones to map their input into a set of features that is identical across the clones. It additionally encourages feature independence and, optionally, reconstruction of a desired target signal by a decoder. As an application, we train a system that extracts a time-sequence of feature vectors of speech and uses it as a conditioning of a WaveNet generative system, facilitating both coding and enhancement.

Streaming spatial audio over networks requires efficient encoding techniques that compress the raw audio content without compromising quality of experience. Streaming service providers such as YouTube need a perceptually relevant objective audio quality metric to monitor users' perceived quality and spatial localization accuracy. In this paper we introduce a full reference objective spatial audio quality metric, AMBIQUAL, which assesses both Listening Quality and Localization Accuracy. In our solution both metrics are derived directly from the B-format Ambisonic audio. The metric extends and adapts the algorithm used in ViSQOLAudio, a full reference objective metric designed for assessing speech and audio quality. In particular, Listening Quality is derived from the omnidirectional channel and Localization Accuracy is derived from a weighted sum of similarity from B-format directional channels. This paper evaluates whether the proposed AMBIQUAL objective spatial audio quality metric can predict two factors: Listening Quality and Localization Accuracy by comparing its predictions with results from MUSHRA subjective listening tests. In particular, we evaluated the Listening Quality and Localization Accuracy of First and Third-Order Ambisonic audio compressed with the OPUS 1.2 codec at various bitrates (i.e. 32, 128 and 256, 512kbps respectively). The sample set for the tests comprised both recorded and synthetic audio clips with a wide range of time-frequency characteristics. To evaluate Localization Accuracy of compressed audio a number of fixed and dynamic (moving vertically and horizontally) source positions were selected for the test samples. Results showed a strong correlation (PCC=0.919; Spearman=0.882 regarding Listening Quality and PCC=0.854; Spearman=0.842 regarding Localization Accuracy) between objective quality scores derived from the B-format Ambisonic audio using AMBIQUAL and subjective scores obtained during listening MUSHRA tests. AMBIQUAL displays very promising quality assessment predictions for spatial audio. Future work will optimise the algorithm to generalise and validate it for any Higher Order Ambisonic formats.