Many real-life applications of automatic speech recognition (ASR) require processing of overlapped speech. A common method involves first separating the speech into overlap-free streams on which ASR is performed. Recently, TF-GridNet has shown impressive performance in speech separation in real reverberant conditions. Furthermore, a mixture encoder was proposed that leverages the mixed speech to mitigate the effect of separation artifacts. In this work, we extended the mixture encoder from a static two-speaker scenario to a natural meeting context featuring an arbitrary number of speakers and varying degrees of overlap. We further demonstrate its limits by the integration with separators of varying strength including TF-GridNet. Our experiments result in a new state-of-the-art performance on LibriCSS using a single microphone. They show that TF-GridNet largely closes the gap between previous methods and oracle separation independent of mixture encoding. We further investigate the remaining potential for improvement.

We propose a general framework to compute the word error rate (WER) of ASR systems that process recordings containing multiple speakers at their input and that produce multiple output word sequences (MIMO). Such ASR systems are typically required, e.g., for meeting transcription. We provide an efficient implementation based on a dynamic programming search in a multi-dimensional Levenshtein distance tensor under the constraint that a reference utterance must be matched consistently with one hypothesis output. This also results in an efficient implementation of the ORC WER which previously suffered from exponential complexity. We give an overview of commonly used WER definitions for multi-speaker scenarios and show that they are specializations of the above MIMO WER tuned to particular application scenarios. We conclude with a discussion of the pros and cons of the various WER definitions and a recommendation when to use which.

Continuous Speech Separation (CSS) has been proposed to address speech overlaps during the analysis of realistic meeting-like conversations by eliminating any overlaps before further processing. CSS separates a recording of arbitrarily many speakers into a small number of overlap-free output channels, where each output channel may contain speech of multiple speakers. Often, a separation model is trained with Utterance-level Permutation Invariant Training (uPIT), which exclusively maps a speaker to an output channel, and applied in a sliding window approach called stitching. Recently, we introduced an alternative training scheme called Graph-PIT that teaches the separator to produce a speaker-shared output channel format without stitching. It can handle an arbitrary number of speakers as long as the number of overlapping speakers is never larger than the number of output channels. Models trained in this way are able to perform segment-less CSS, i.e., without stitching, and achieve comparable and often better separation quality than the conventional CSS with uPIT and stitching. In this contribution, we further investigate the Graph-PIT training scheme. We show in extended experiments that Graph-PIT also works in challenging reverberant conditions. We simplify the training schedule for Graph-PIT with the recently proposed Source Aggregated Signal-to-Distortion Ratio (SA-SDR) loss, which eliminates unfavorable properties of the previously used A-SDR loss to enable training with Graph-PIT from scratch. Furthermore, we introduce novel signal-level evaluation metrics for meeting scenarios, namely the source-aggregated scale- and convolution-invariant Signal-to-Distortion Ratio (SA-SI-SDR and SA-CI-SDR), which are generalizations of the commonly used SDR-based metrics for the CSS case.

Many state-of-the-art neural network-based source separation systems use the averaged Signal-to-Distortion Ratio (SDR) as a training objective function. The basic SDR is, however, undefined if the network reconstructs the reference signal perfectly or if the reference signal contains silence, e.g., when a two-output separator processes a single-speaker recording. Many modifications to the plain SDR have been proposed that trade-off between making the loss more robust and distorting its value. We propose to switch from a mean over the SDRs of each individual output channel to a global SDR over all output channels at the same time, which we call source-aggregated SDR (SA-SDR). This makes the loss robust against silence and perfect reconstruction as long as at least one reference signal is not silent. We experimentally show that our proposed SA-SDR is more stable and preferable over other well-known modifications when processing meeting-style data that typically contains many silent or single-speaker regions.

Recent speaker diarization studies showed that integration of end-to-end neural diarization (EEND) and clustering-based diarization is a promising approach for achieving state-of-the-art performance on various tasks.Such an approach first divides an observed signal into fixed-length segments, then performs segment-level local diarization based on an EEND module, and merges the segment-level results via clustering to form a final global diarization result.The segmentation is done to limit the number of speakers in each segment since the current EEND cannot handle a large number of speakers.In this paper, we argue that such an approach involving the segmentation has several issues; for example, it inevitably faces a dilemma that larger segment sizes increase both the context available for enhancing the performance and the number of speakers for the local EEND module to handle.To resolve such a problem, this paper proposes a novel framework that performs diarization without segmentation.However, it can still handle challenging data containing many speakers and a significant amount of overlapping speech.The proposed method can take an entire meeting for inference and perform utterance-by-utterance diarization that clusters utterance activities in terms of speakers.To this end, we leverage a neural network training scheme called Graph-PIT proposed recently for neural source separation.Experiments with simulated active-meeting-like data and CALLHOME data show the superiority of the proposed approach over the conventional methods.

Automatic transcription of meetings requires handling of overlapped speech, which calls for continuous speech separation (CSS) systems. The uPIT criterion was proposed for utterance-level separation with neural networks and introduces the constraint that the total number of speakers must not exceed the number of output channels. When processing meeting-like data in a segment-wise manner, i.e., by separating overlapping segments independently and stitching adjacent segments to continuous output streams, this constraint has to be fulfilled for any segment. In this contribution, we show that this constraint can be significantly relaxed. We propose a novel graph-based PIT criterion, which casts the assignment of utterances to output channels in a graph coloring problem. It only requires that the number of concurrently active speakers must not exceed the number of output channels. As a consequence, the system can process an arbitrary number of speakers and arbitrarily long segments and thus can handle more diverse scenarios. Further, the stitching algorithm for obtaining a consistent output order in neighboring segments is of less importance and can even be eliminated completely, not the least reducing the computational effort. Experiments on meeting-style WSJ data show improvements in recognition performance over using the uPIT criterion.

The rising interest in single-channel multi-speaker speech separation sparked development of End-to-End (E2E) approaches to multi-speaker speech recognition. However, up until now, state-of-the-art neural network-based time domain source separation has not yet been combined with E2E speech recognition. We here demonstrate how to combine a separation module based on a Convolutional Time domain Audio Separation Network (Conv-TasNet) with an E2E speech recognizer and how to train such a model jointly by distributing it over multiple GPUs or by approximating truncated back-propagation for the convolutional front-end. To put this work into perspective and illustrate the complexity of the design space, we provide a compact overview of single-channel multi-speaker recognition systems. Our experiments show a word error rate of 11.0% on WSJ0-2mix and indicate that our joint time domain model can yield substantial improvements over cascade DNN-HMM and monolithic E2E frequency domain systems proposed so far.

Recently, the source separation performance was greatly improved by time-domain audio source separation based on dualpath recurrent neural network (DPRNN).DPRNN is a simple but effective model for a long sequential data.While DPRNN is quite efficient in modeling a sequential data of the length of an utterance, i.e., about 5 to 10 second data, it is harder to apply it to longer sequences such as whole conversations consisting of multiple utterances.It is simply because, in such a case, the number of time steps consumed by its internal module called inter-chunk RNN becomes extremely large.To mitigate this problem, this paper proposes a multi-path RNN (MPRNN), a generalized version of DPRNN, that models the input data in a hierarchical manner.In the MPRNN framework, the input data is represented at several (≥ 3) time-resolutions, each of which is modeled by a specific RNN sub-module.For example, the RNN sub-module that deals with the finest resolution may model temporal relationship only within a phoneme, while the RNN sub-module handling the most coarse resolution may capture only the relationship between utterances such as speaker information.We perform experiments using simulated dialogue-like mixtures and show that MPRNN has greater model capacity, and it outperforms the current state-of-the-art DPRNN framework especially in online processing scenarios.

Most approaches to multi-talker overlapped speech separation and recognition assume that the number of simultaneously active speakers is given, but in realistic situations, it is typically unknown. To cope with this, we extend an iterative speech extraction system with mechanisms to count the number of sources and combine it with a single-talker speech recognizer to form the first end-to-end multi-talker automatic speech recognition system for an unknown number of active speakers. Our experiments show very promising performance in counting accuracy, source separation and speech recognition on simulated clean mixtures from WSJ0-2mix and WSJ0-3mix. Among others, we set a new state-of-the-art word error rate on the WSJ0-2mix database. Furthermore, our system generalizes well to a larger number of speakers than it ever saw during training, as shown in experiments with the WSJ0-4mix database.

Automatic meeting analysis comprises the tasks of speaker counting, speaker diarization, and the separation of overlapped speech, followed by automatic speech recognition. This all has to be carried out on arbitrarily long sessions and, ideally, in an online or block-online manner. While significant progress has been made on individual tasks, this paper presents for the first time an all-neural approach to simultaneous speaker counting, diarization and source separation. The NN-based estimator operates in a block-online fashion and tracks speakers even if they remain silent for a number of time blocks, thus learning a stable output order for the separated sources. The neural network is recurrent over time as well as over the number of sources. The simulation experiments show that state of the art separation performance is achieved, while at the same time delivering good diarization and source counting results. It even generalizes well to an unseen large number of blocks.

Deep clustering (DC) and deep attractor networks (DANs) are a data-driven way to monaural blind source separation. Both approaches provide astonishing single channel performance but have not yet been generalized to block-online processing. When separating speech in a continuous stream with a block-online algorithm, it needs to be determined in each block which of the output streams belongs to whom. In this contribution we solve this block permutation problem by introducing an additional speaker identification embedding to the DAN model structure. We motivate this model decision by analyzing the embedding topology of DC and DANs and show, that DC and DANs themselves are not sufficient for speaker identification. This model structure (a) improves the signal to distortion ratio (SDR) over a DAN baseline and (b) provides up to 61% and up to 34% relative reduction in permutation error rate and re-identification error rate compared to an i-vector baseline, respectively.