In this paper, we describe SpeakerStew - a hybrid system to perform speaker verification on 46 languages. Two core ideas were explored in this system: (1) Pooling training data of different languages together for multilingual generalization and reducing development cycles; (2) A triage mechanism between text-dependent and text-independent models to reduce runtime cost and expected latency. To the best of our knowledge, this is the first study of speaker verification systems at the scale of 46 languages. The problem is framed from the perspective of using a smart speaker device with interactions consisting of a wake-up keyword (text-dependent) followed by a speech query (text-independent).Experimental evidence suggests that training on multiple languages can generalize to unseen varieties while maintaining performance on seen varieties. We also found that it can reduce computational requirements for training models by an order of magnitude. Furthermore, during model inference on English data, we observe that leveraging a triage framework can reduce the number of calls to the more computationally expensive text-independent system by 73% (and reduce latency by 60%) while maintaining an EER no worse than the text-independent setup.
In this work we study some of the challenges associated with scaling speaker recognition systems to multiple languages. To the best of our knowledge, this is the first study of speaker verification systems at the scale of 46 languages. Training models for each of the many languages can be time and energy demanding in addition to costly. Low resource languages present additional difficulties. The problem is framed from the perspective of using a smart speaker device with interactions consisting of a wake-up keyword (text-dependent) followed by a speech query (text-independent). We examine the use of a hybrid setup consisting of multilingual text-dependent and text-independent components. Experimental evidence suggests that training on multiple languages can generalize to unseen varieties while maintaining performance on seen varieties. We also found that it can reduce computational requirements for training models by an order of magnitude. Furthermore, during model inference on English data, we observe that leveraging a triage framework can reduce the number of calls to the more computationally expensive text-independent system by 73% (and reduce latency by 60%) while maintaining an EER no worse than the text-independent setup.
Many neural network speaker recognition systems model each speaker using a fixed-dimensional embedding vector. These embeddings are generally compared using either linear or 2nd-order scoring and, until recently, do not handle utterance-specific uncertainty. In this work we propose scoring these representations in a way that can capture uncertainty, enroll/test asymmetry and additional non-linear information. This is achieved by incorporating a 2nd-stage neural network (known as a decision network) as part of an end-to-end training regimen. In particular, we propose the concept of decision residual networks which involves the use of a compact decision network to leverage cosine scores and to model the residual signal that's needed. Additionally, we present a modification to the generalized end-to-end softmax loss function to better target the separation of same/different speaker scores. We observed significant performance gains for the two techniques.
We introduce VoiceFilter-Lite, a single-channel source separation model that runs on the device to preserve only the speech signals from a target user, as part of a streaming speech recognition system. Delivering such a model presents numerous challenges: It should improve the performance when the input signal consists of overlapped speech, and must not hurt the speech recognition performance under all other acoustic conditions. Besides, this model must be tiny, fast, and perform inference in a streaming fashion, in order to have minimal impact on CPU, memory, battery and latency. We propose novel techniques to meet these multi-faceted requirements, including using a new asymmetric loss, and adopting adaptive runtime suppression strength. We also show that such a model can be quantized as a 8-bit integer model and run in realtime.
We demonstrate that a production-quality keyword-spotting model can be trained on-device using federated learning and achieve comparable false accept and false reject rates to a centrally-trained model. To overcome the algorithmic constraints associated with fitting on-device data (which are inherently non-independent and identically distributed), we conduct thorough empirical studies of optimization algorithms and hyperparameter configurations using large-scale federated simulations. To overcome resource constraints, we replace memory intensive MTR data augmentation with SpecAugment, which reduces the false reject rate by 56%. Finally, to label examples (given the zero visibility into on-device data), we explore teacher-student training.
Google's multilingual speech recognition system combines low-level acoustic signals with language-specific recognizer signals to better predict the language of an utterance. This paper presents our experience with different signal combination methods to improve overall language identification accuracy. We compare the performance of a lattice-based ensemble model and a deep neural network model to combine signals from recognizers with that of a baseline that only uses low-level acoustic signals. Experimental results show that the deep neural network model outperforms the lattice-based ensemble model, and it reduced the error rate from 5.5% in the baseline to 4.3%, which is a 21.8% relative reduction.
In this paper, we propose "personal VAD", a system to detect the voice activity of a target speaker at the frame level. This system is useful for gating the inputs to a streaming speech recognition system, such that it only triggers for the target user, which helps reduce the computational cost and battery consumption. We achieve this by training a VAD-alike neural network that is conditioned on the target speaker embedding or the speaker verification score. For every frame, personal VAD outputs the scores for three classes: non-speech, target speaker speech, and non-target speaker speech. With our optimal setup, we are able to train a 130KB model that outperforms a baseline system where individually trained standard VAD and speaker recognition network are combined to perform the same task.
In many scenarios of a language identification task, the user will specify a small set of languages which he/she can speak instead of a large set of all possible languages. We want to model such prior knowledge into the way we train our neural networks, by replacing the commonly used softmax loss function with a novel loss function named tuplemax loss. As a matter of fact, a typical language identification system launched in North America has about 95% users who could speak no more than two languages. Using the tuplemax loss, our system achieved a 2.33% error rate, which is a relative 39.4% improvement over the 3.85% error rate of standard softmax loss method.