Many modern systems for speaker diarization, such as the recently-developed VBx approach, rely on clustering of DNN speaker embeddings followed by resegmentation.Two problems with this approach are that the DNN is not directly optimized for this task, and the parameters need significant retuning for different applications.We have recently presented progress in this direction with a Leave-One-Out Gaussian PLDA (LGP) clustering algorithm and an approach to training the DNN such that embeddings directly optimize performance of this scoring method.This paper presents a new two-pass version of this system, where the second pass uses finer time resolution to significantly improve overall performance.For the Callhome corpus, we achieve the first published error rate below 4% without any task-dependent parameter tuning.We also show significant progress towards a robust single solution for multiple diarization tasks.

The VoxSRC speaker recognition challenge comprises data obtained from YouTube videos of celebrity interviews in a wide range of recording environments. The challenge provides FIXED and OPEN training conditions to allow cross-system comparisons and to characterize the effects of additional amounts of training data on system performance. This paper describes our submission to this challenge where we have explored x-vector extractor topologies, classification head alternatives, data augmentation, and angular margin penalty. Our final entry to the FIXED condition (which achieved 2nd place) is the score average of 4 diverse systems. We find that this system outperforms a large single DNN with similar number of parameters.

The effects of speaking-style variability on automatic speaker verification were investigated using the UCLA Speaker Variability database which comprises multiple speaking styles per speaker. An x-vector/PLDA (probabilistic linear discriminant analysis) system was trained with the SRE and Switchboard databases with standard augmentation techniques and evaluated with utterances from the UCLA database. The equal error rate (EER) was low when enrollment and test utterances were of the same style (e.g., 0.98% and 0.57% for read and conversational speech, respectively), but it increased substantially when styles were mismatched between enrollment and test utterances. For instance, when enrolled with conversation utterances, the EER increased to 3.03%, 2.96% and 22.12% when tested on read, narrative, and pet-directed speech, respectively. To reduce the effect of style mismatch, we propose an entropy-based variable frame rate technique to artificially generate style-normalized representations for PLDA adaptation. The proposed system significantly improved performance. In the aforementioned conditions, the EERs improved to 2.69% (conversation -- read), 2.27% (conversation -- narrative), and 18.75% (pet-directed -- read). Overall, the proposed technique performed comparably to multi-style PLDA adaptation without the need for training data in different speaking styles per speaker.

Recently, deep neural networks that map utterances to fixed-dimensional embeddings have emerged as the state-of-the-art in speaker recognition. Our prior work introduced x-vectors, an embedding that is very effective for both speaker recognition and diarization. This paper combines our previous work and applies it to the problem of speaker recognition on multi-speaker conversations. We measure performance on Speakers in the Wild and report what we believe are the best published error rates on this dataset. Moreover, we find that diarization substantially reduces error rate when there are multiple speakers, while maintaining excellent performance on single-speaker recordings. Finally, we introduce an easily implemented method to remove the domain-sensitive threshold typically used in the clustering stage of a diarization system. The proposed method is more robust to domain shifts, and achieves similar results to those obtained using a well-tuned threshold.

In this paper, we apply several modifications to script identification, several of which inspired by techniques from the similar audio task of spoken language recognition. Specifically, we alter the architecture of a convolutional network with global average pooling to include variance pooling as well, we utilize score calibration of the output scores of the network, and we utilize prior distribution estimation to condition the calibrated scores. We show that these methods are effective in script identification, with the use of priors showing especially promising improvements. Furthermore, in the domain of script identification, several additional extensions of distribution estimation are available which consider the distribution within each image, and we demonstrate much larger improvements when employing these extensions. Finally, we also show that an embedding-plus-classifier approach performs similarly to the full network, and so its potential for increased flexibility may be beneficial for future consideration. With all modifications, overall accuracy on the ICDAR 2017 validation dataset increases from 89.7% to 93.6%.

Speaker diarization is an important front-end for many speech technologies in the presence of multiple speakers, but current methods that employ i-vector clustering for short segments of speech are potentially too cumbersome and costly for the front-end role. In this work, we propose an alternative approach for learning representations via deep neural networks to remove the i-vector extraction process from the pipeline entirely. The proposed architecture simultaneously learns a fixed-dimensional embedding for acoustic segments of variable length and a scoring function for measuring the likelihood that the segments originated from the same or different speakers. Through tests on the CALLHOME conversational telephone speech corpus, we demonstrate that, in addition to streamlining the diarization architecture, the proposed system matches or exceeds the performance of state-of-the-art baselines. We also show that, though this approach does not respond as well to unsupervised calibration strategies as previous systems, the incorporation of well-founded speaker priors sufficiently mitigates this shortcoming.

I-vector training and extraction assume that a speech file is spoken by a single speaker. This work considers the effects of violating that assumption with the presence of cross-talk or multi-speaker conversations. First, it is demonstrated that these problematic speech files can be detected using the i-vector representation itself. The impact of these violations of the single-speaker assumption are then explored along with strategies to mitigate it. It is shown that, even in predominantly clean data, the removal of cross-talk can provide modest gains, but that T matrix and PLDA training are largely robust to these types of noise. It is also shown that detection in front of diarization is a reasonable strategy in the presence of data with an unknown proportion of multi-speaker conversations. Finally, in the course of this work, evidence is found that cross-talk detection and multi-speaker detection may in fact be different tasks that require separately trained detectors.

We compare the multinomial i-vector framework from the speech community with LDA, SAGE, and LSA as feature learners for topic ID on multinomial speech and text data.We also compare the learned representations in their ability to discover topics, quantified by distributional similarity to gold-standard topics and by human interpretability.We find that topic ID and topic discovery are competing objectives.We argue that LSA and i-vectors should be more widely considered by the text processing community as pre-processing steps for downstream tasks, and also speculate about speech processing tasks that could benefit from more interpretable representations like SAGE.

Content-based recommender systems use preference ratings and features that characterize media to model users' interests or information needs for making future recommendations. While previously developed in the music and text domains, we present an initial exploration of content-based recommendation for spoken documents using a corpus of public domain internet audio. Unlike familiar speech technologies of topic identification and spoken document retrieval, our recommendation task requires a more comprehensive notion of document relevance than bags-of-words would supply. Inspired by music recommender systems, we automatically extract a wide variety of content-based features to characterize non-linguistic aspects of the audio such as speaker, language, gender, and environment. To combine these heterogeneous information sources into a single relevance judgement, we evaluate feature, score, and hybrid fusion techniques. Our study provides an essential first exploration of the task and clearly demonstrates the value of a multisource approach over a bag-of-words baseline.

In this paper, we present a comprehensive study on supervised domain adaptation of PLDA based i-vector speaker recognition systems. After describing the system parameters subject to adaptation, we study the impact of their adaptation on recognition performance. Using the recently designed domain adaptation challenge, we observe that the adaptation of the PLDA parameters (i.e. across-class and within-class co variances) produces the largest gains. Nonetheless, length-normalization is also important; whereas using an indomani UBM and T matrix is not crucial. For the PLDA adaptation, we compare four approaches. Three of them are proposed in this work, and a fourth one was previously published. Overall, the four techniques are successful at leveraging varying amounts of labeled in-domain data and their performance is quite similar. However, our approaches are less involved, and two of them are applicable to a larger class of models (low-rank across-class).

During late-2013 through early-2014 NIST coordinated a special i-vector challenge based on data used in previous NIST Speaker Recognition Evaluations (SREs). Unlike evaluations in the SRE series, the i-vector challenge was run entirely online and used fixed-length feature vectors projected into a low-dimensional space (i-vectors) rather than audio recordings. These changes made the challenge more readily accessible, especially to participants from outside the audio processing field. Compared to the 2012 SRE, the i-vector challenge saw an increase in the number of participants by nearly a factor of two, and a two orders of magnitude increase in the number of systems submitted for evaluation. Initial results indicate the leading system achieved an approximate 37% improvement relative to the baseline system.

Traditional i-vector speaker recognition systems use a Gaussian mixture model (GMM) to collect sufficient statistics (SS). Recently, replacing this GMM with a deep neural network (DNN) has shown promising results. In this paper, we explore the use of DNNs to collect SS for the unsupervised domain adaptation task of the Domain Adaptation Challenge (DAC).We show that collecting SS with a DNN trained on out-of-domain data boosts the speaker recognition performance of an out-of-domain system by more than 25%. Moreover, we integrate the DNN in an unsupervised adaptation framework, that uses agglomerative hierarchical clustering with a stopping criterion based on unsupervised calibration, and show that the initial gains of the out-of-domain system carry over to the final adapted system. Despite the fact that the DNN is trained on the out-of-domain data, the final adapted system produces a relative improvement of more than 30% with respect to the best published results on this task.

This paper presents a framework for fully Bayesian speaker comparison of i-vectors. By generalizing the train/test paradigm, we derive an analytic expression for the speaker comparison log-likelihood ratio (LLR), as well as solutions for model training and Bayesian scoring. This framework is useful for enrollment sets of any size. For the specific case of single-cut enrollment, it is shown to be mathematically equivalent to probabilistic linear discriminant analysis (PLDA). Additionally, we present discriminative training of model hyper-parameters by minimizing the total cross entropy between LLRs and class labels. When applied to speaker recognition, significant performance gains are observed for various NIST SRE 2010 extended evaluation tasks.

In this paper we propose fully Bayesian speaker comparison of supervectors, which we refer to as SV-BSC, as a method for estimating whether a test cut was generated by the same speaker as an enrollment set. We derive the SV-BSC log-likelihood ratio of same-speaker to different-speaker hypotheses, and present solutions for model training and Bayesian scoring. We then show that if speaker and channel variability are assumed to inhabit a total variability subspace, SV-BSC scoring reduces to a form which requires only low-computation subspace operations. Finally, we show that common speaker recognition techniques such as Joint Factor Analysis (JFA) and i-vector Probabilistic Linear Discriminant Analysis (PLDA) are approximations to this full solution under certain additional assumptions. Experiments on the NIST 2010 SRE show SV-BSC to outperform a PLDA system.

We propose a method for spectral enhancement of reverberant speech based on inverting the modulation transfer function (MTF). Using all-pole models of modulation spectra allows the linear prediction inverse MTF (LP-IMTF) filter to exhibit a smooth frequency response, and allows it to be implemented as a low-order IIR filter in the modulation envelope domain. The proposed filter adapts to current acoustic conditions without relying on explicit information regarding reverberation time. Additionally, the LP-IMTF framework allows for estimation of useful side information, such as local signal-to-reverberation ratios and band-specific reverberation times. As example applications, the LP-IMTF system is applied to enhancement and speaker recognition of reverberant speech, and significant performance improvements are achieved.

Research in the speaker recognition community has continued to ad dress methods of mitigating variational nuisances. Telephone and auxiliary-microphone recorded speech emphasize the need for a ro bust way of dealing with unwanted variation. The design of recent 2010 NIST-SRE Speaker Recognition Evaluation (SRE) reflects this research emphasis. In this paper, we present the MIT submission applied to the tasks of the 2010 NIST-SRE with two main goals- language-independent scalable modeling and robust nuisance mitigation. For modeling, exclusive use of inner product-based and cepstral systems produced a language-independent computationally scalable system. For robustness, systems that captured spectral and prosodic information, modeled nuisance subspaces using multiple novel methods, and fused scores of multiple systems were implemented. The performance of the system is presented on a subset of the NIST SRE 2010 core tasks.

Abstract : We present a new perspective on the subspace compensation techniques that currently dominate the field of speaker recognition using Gaussian Mixture Models (GMMs). Rather than the traditional factor analysis approach, we use Gaussian modeling in the sufficient statistic supervector space combined with Probabilistic Principal Component Analysis (PPCA) within-class and shared across class covariance matrices to derive a family of training and testing algorithms. Key to this analysis is the use of two noise terms for each speech cut: a random channel offset and a length dependent observation noise. Using the Wiener filtering perspective, formulas for optimal train and test algorithms for Joint Factor Analysis (JFA) are simple to derive. In addition, we can show that an alternative form of Wiener filtering results in the i-vector approach. thus tying together these two disparate techniques.

Abstract 1 Of increasing importance in the civilian and military population is the recognition of Major Depressive Disorder at its earliest stages and intervention before the onset of severe symptoms. Toward the goal of more effective monitoring of depression severity, we investigate automatic classifiers of depression state, that have the important property of mitigating nuisances due to data variability, such as speaker and channel effects, unrelated to levels of depression. To assess our measures, we use a 35-speaker free-response speech database of subjects treated for depression over a six-week duration, along with standard clinical HAMD depression ratings. Preliminary experiments indicate that by mitigating nuisances, thus focusing on depression severity as a class, we can significantly improve classification accuracy over baseline Gaussian-mixture-model-based classifiers. Index Terms: major depressive disorder, Gaussian-mixture models, joint factor analysis, nuisance mitigation

This paper presents a description of the MIT Lincoln Laboratory language recognition system submitted to the NIST 2009 Language Recognition Evaluation (LRE). This system consists of a fusion of three core recognizers, two based on spectral similarity and one based on tokenization. The 2009 LRE differed from previous ones in that test data included narrowband segments from worldwide Voice of America broadcasts as well as conventional recorded conversational telephone speech. Results are presented for the 23-language closed-set and open-set detection tasks at the 30, 10, and 3 second durations along with a discussion of the language-pair task. On the 30 second 23-language closed set detection task, the system achieved a 1.64 average error rate.

We present two approaches to noise robust very low bit rate speech coding using wideband MELP analysis/synthesis. Both methods exploit multiple acoustic and non-acoustic input sensors, using our previously-presented dynamic waveform fusion algorithm to simultaneously perform waveform fusion, noise suppression, and cross-channel noise cancellation. One coder uses a 600 bps scalable phonetic vocoder, with a phonetic speech recognizer followed by joint predictive vector quantization of the error in wideband MELP parameters. The second coder operates at 300 bps with fixed 80 ms segments, using novel variable-rate multistage matrix quantization techniques. Formal test results show that both coders achieve equivalent intelligibility to the 2.4 kbps NATO standard MELPe coder in harsh acoustic noise environments, at much lower bit rates, with only modest quality loss.

This paper presents a description of the MIT Lincoln Laboratory language recognition system submitted to the NIST 2007 Language Recognition Evaluation. This system consists of a fusion of four core recognizers, two based on tokenization and two based on spectral similarity. Results for NIST’s 14-language detection task are presented for both the closed-set and open-set tasks and for the 30, 10 and 3 second durations. On the 30 second 14-language closed set detection task, the system achieves a 1% equal error rate.

Speech communication is significantly more difficult in severe acoustic background noise environments, especially when low-rate speech coders are used. Non-acoustic sensors, such as radar sensors, vibrometers, and bone-conduction microphones, offer significant potential in these situations. We extend previous work on fixed waveform fusion from multiple sensors to an optimal dynamic waveform fusion algorithm that minimizes both additive noise and signal distortion in the estimated speech signal. We show that a minimum mean squared error (MMSE) waveform matching criterion results in a generalized multichannel Wiener filter, and that this filter will simultaneously perform waveform fusion, noise suppression, and crosschannel noise cancellation. Formal intelligibility and quality testing demonstrate significant improvement from this approach.

We present the framework for a scalable phonetic vocoder (SPV) capable of operating at bit rates from 300 - 1100 bps. The underlying system uses an HMM-based phonetic speech recognizer to estimate the parameters for MELP speech synthesis. We extend this baseline technique in three ways. First, we introduce the concept of predictive time evolution to generate a smoother path for the synthesizer parameters, and show that it improves speech quality. Then, since the output speech from the phonetic vocoder is still limited by such low bit rates, we propose a scalable system where the accuracy of the MELP parameters is increased by vector quantizing the error signal between the true and phonetic-estimated MELP parameters. Finally, we apply an extremely flexible technique for exploiting correlations in these parameters over time, which we call joint predictive vector quantization (JPVQ). We show that significant quality improvement can be attained by adding as few as 400 bps to the baseline phonetic vocoder using JPVQ. The resulting SPV system provides a flexible platform for adjusting the phonetic vocoder bit rate and speech quality

In this paper, we investigate the degradation of speaker identification performance due to speech coding algorithms used in digital telephone networks, cellular telephony, and voice over IP. By analyzing the difference between front-end feature vectors derived from coded and uncoded speech in terms of spectral distortion, we are able to quantify this coding degradation. This leads to two novel methods for distortion compensation: codebook and LPC compensation. Both are shown to significantly reduce front-end mismatch, with the second approach providing the most encouraging results. Full experiments using a GMM-UBM speaker ID system confirm the usefulness of both the front-end distortion analysis and the LPC compensation technique. 1.

It is well-known that wideband speech (0-7 kHz) provides better quality and intelligibility than narrowband speech (300-3400 Hz), but typically only narrowband speech information is available in current wireless communication systems. Narrowband to wideband extension technology has been recently investigated to artificially generate wideband speech from narrowband speech for better speech quality and intelligibility. This paper presents a robust split-band narrowband to wideband extension system based on algorithmic enhancements to the codebook mapping technique for high-band parameter estimation. Numerical measurements confirm the performance improvements of the codebook mapping process, and informal listening evaluations show the potential of the system and its robustness to input distortions and non-speech input signals.

This paper describes extensions of the 4 kb/s hybrid MELP/CELP coder, up to 6.4 kb/s and down to 2.4 kb/s. The baseline 4 kb/s coder uses three coding modes: MELP in strongly voiced speech frames, CELP with pitch prediction in weakly voiced frames, and CELP with stochastic excitation in unvoiced frames. To minimize switching artifacts between parametric MELP and waveform CELP coding, an alignment phase is encoded in MELP and zero-phase equalization is applied to the CELP target signal. The 6.4 kb/s extension uses the same three modes as the 4 kb/s coder, with improved MELP and CELP coders. The 2.4 kb/s extension uses only two modes: MELP for voiced frames and CELP synthesis with random excitation for unvoiced frames. The alignment phase is encoded in MELP frames for all bit rates so that time synchrony with input speech is always maintained. Alignment phase and zero-phase equalization enable smooth switching between coders at different bit rates. The hybrid MELP/CELP coding structure leads to coders that perform better at a given bit rate than MELP or CELP separately, and better than or equivalent to higher bit-rate ITU standards. Formal subjective tests show that for all-but-one tested conditions, the 6.4 kb/s hybrid coder is better than 8 kb/s G.729 and the 2.4 kb/s coder is equivalent to, or better than, 6.4 kb/s G.729 Annex D.

This paper presents an improved 4 kb/s hybrid MELP/CELP speech coder submitted as a candidate for ITU standardization. The coder uses three modes: a high-quality MELP coder for strongly voiced speech frames, an ACELP coder with pitch prediction for weakly voiced frames, and a stochastic CELP coder for unvoiced frames. We present recent enhancements to this coder, both to improve speech quality and to reduce coder complexity. Previous lTU Selection Testing results on an earlier version of this coder showed that it met nearly all requirements for toll-quality speech, more than any other candidate. Our internal testing shows that the current reduced-complexity fixed-point coder maintains this high performance.

We propose an adaptive voice playout method for handling network delay jitter in voice over packet (VOP) receivers. Our method allows playout delay increase during both silence periods and active speech; however, it allows playout delay decrease during silence periods only. Since the playout delay is increased during active speech, a speech extension algorithm is required. Therefore we have developed a novel, voicing-classification based speech extension algorithm for CELP speech coders. Though the complementary speech truncation algorithm is not needed for our adaptive playout mechanism, it is suitable for applications such as voice synchronization with other media.

This paper presents a multi-rate wideband speech coder with bit rates from 8 to 32 kb/s. The coder uses a splitband approach, where the input signal, sampled at 16 kHz, is split into two equal frequency bands from 0-4 kHz and 4-8 kHz, each of which is decimated to an 8 kHz sampling rate. The lower band is coded using the adaptive multi-rate (AMR) family of high-quality narrowband speech coders, while the higher band is represented by a simple but effective parametric model. A complete solution including this wideband speech coder, channel coding for various GSM channels, and dynamic rate adaptation, easily passed all Selection Rules and ranked second overall in the 3GPP AMR Wideband Selection Testing. Besides the high performance, additional advantages of the embedded split-band approach include ease of implementation, reduced complexity, and simplified interoperation with narrowband speech coders.

Diversity schemes include information about packet n in future packets or send information about packet n via separate paths. If packet n is lost, it is reconstructed from information included in future packets or information received via separate paths. This paper presents CELP-based diversity schemes for voice over packet applications. The diversity schemes reduce the impact of packet losses while being efficient in terms of both bandwidth requirement and computational complexity. With our diversity schemes, transmission schemes that allocate bandwidth resources among diversity stages during congestion give significantly better performance than schemes that use no diversity during congestion, for the same bandwidth usage.

This paper describes a new 14 kb/s wideband speech coder. The coder uses a split-band approach, where the input signal, sampled at 16 kHz, is split into two equal frequency bands from 0-4 kHz and 4-8 kHz, each of which is decimated to an 8 kHz sampling rate. The lower band is coded with a high-quality narrowband speech coder, the 11.8 kb/s G.729 Annex E, while the higher band is represented by a simple but effective parametric model. Two new features facilitate efficient coding of the high-band signal: noise modulation and high-frequency reversal. Since the encoding of the lower band is independent of the high-band signal, the narrowband encoder output can be embedded in the overall bitstream. Subjective test results show that this wideband speech coder is capable of producing high quality output speech.

The paper describes a hybrid multi-modal codec with MELP and CELP coders used for different speech regions. Three modes are used: strongly voiced, weakly voiced, and unvoiced. The weakly voiced mode includes transitions and plosives; it is used when neither strong voicing nor unvoiced region are clearly identified. In the strongly voiced mode the MELP coder is used, while in the weakly voiced and unvoiced modes the CELP coder is employed. To limit switching artifacts between the coders, alignment phase is estimated and transmitted in the MELP mode making the original and MELP-synthesized speech time-synchronous. Additionally, in zero-phase equalization, the phase component of the CELP target signal is removed making the target waveform more similar to the MELP-synthesized speech. These two techniques, alignment-phase encoding and zero-phase equalization, greatly reduce switching artifacts in MELP/CELP transition regions. Formal listening test results of the 4 kb/s hybrid coder show that it can achieve speech quality equivalent to 32 kb/s ADPCM.

We have developed an adaptive multi-rate (AMR) speech coder designed to operate under the GSM digital cellular full rate (22.8 kb/s) and half rate (11.4 kb/s) channels and to maintain high quality in the presence of highly varying background noise and channel conditions. Within each total rate, several codec modes with different source/channel bit rate allocations are used. The speech coders in each codec mode are based on the CELP algorithm operating at rates ranging from 11.85 kb/s down to 5.15 kb/s, where the lowest rate coder is a source controlled multi-modal speech coder. The decoders monitor the channel quality at both ends of the wireless link using the soft values for the received bits and assist the base station in selecting the codec mode that is appropriate for a given channel condition. The coder was submitted to the GSM AMR standardization competition and met the qualification requirements in an independent formal MOS test.

A number of coding techniques have been reported to achieve near toll quality synthesized speech at bit-rates around 4 kb/s. These include variants of code excited linear prediction (CELP), sinusoidal transform coding (STC) and multi-band excitation (MBE). While CELP has been an effective technique for bit-rates above 6 kb/s, STC, MBE, waveform interpolation (WI) and mixed excitation linear prediction (MELP) models seem to be attractive at bit-rates below 3 kb/s. We present a system to encode speech with high quality using MELP, a technique previously demonstrated to be effective at bit-rates of 1.6-2.4 kb/s. We have enhanced the MELP model producing significantly higher speech quality at bit-rates above 2.4 kb/s. We describe the development and testing of a high quality 4 kb/s MELP coder.

It is well-known that an impulse-excited, all-pole filter is capable of representing many physical phenomena, including the oscillatory modes of percussion musical instruments like woodblocks, xylophones, or chimes. In contrast to the more common application of all-pole models to speech, however, practical problems arise in music synthesis due to the location of poles very close to the unit circle. The objective of this work was to develop algorithms to find excitation and filter parameters for synthesis of percussion instrument sounds using only an inexpensive all-pole filter chip (TI TSP50C1x). The paper describes analysis methods for dealing with pole locations near the unit circle, as well as a general method for modeling the transient attack characteristics of a particular sound while independently controlling the amplitudes of each oscillatory mode.

This paper describes our new mixed excitation linear predictive (MELP) coder designed for very low bit rate applications. This new coder, through algorithmic improvements and enhanced quantization techniques, produces better speech quality at 1.7 kb/s than the new U.S. Federal Standard MELP coder at 2.4 kb/s. Key features of the coder are an improved pitch estimation algorithm and a line spectral frequencies (LSF) quantization scheme that requires only 21 bits per frame. With channel coding, this new MELP coder is capable of maintaining good speech quality even in severely degraded channels, at a total bit rate of only 3 kb/s.

In general, a variable rate coder can obtain the same speech quality as a fixed rate coder, while reducing the average bit rate. We have developed a variable-rate multimodal speech coder with an average bit rate of 3 kb/s for a speech activity factor of 80% and quality comparable to the GSM full rate coder. The coder has four coding modes and uses a robust classification method involving the pitch gain, zero crossings, and a peakiness measure. Also the coder employs a novel gain-matched analysis-by-synthesis technique for very low rate coding of unvoiced frames and an improved noise-level-dependent postfilter. This paper describes the details of our algorithm and presents the results from subjective listening tests.

This paper describes the new U.S. Federal Standard at 2400 bps. The mixed excitation linear prediction (MELP) coder was chosen by the DoD Digital Voice Processing Consortium to replace the existing 2400 bps Federal Standard FS 1015 (LPC-10). This new standard provides equal or improved performance over the 4800 bps Federal Standard FS 1016 (CELP) at a rate equivalent to LPC-10. The MELP coder is based on the traditional LPC model, but includes additional features to improve its performance.

Remote retrieval of compressed voicemail data over a telephone line is one of several emerging applications of speech coding. Using a notebook computer equipped with a modem, a user can remotely access a networked desktop unit located at their office or home to retrieve various types of information such as email, FAX, electronic documents as well as voicemail. By compressing the speech data, we reduce the amount of time needed for the transfer of voice data over the telephone line. We have developed a high-quality variable-rate CELP (code-excited linear prediction) coder which can be used in such an application. The coder operates at an average rate of 3 kb/s assuming 80% speech activity and uses several new techniques including a mode decision mechanism based on the use of a peakiness measure and a novel excitation search method called gain-matched analysis-by-synthesis. The coder gives quality comparable to the Microsoft GSM 6.10 Audio Codec at 13 kb/s.

This paper presents a block-coded VFR strategy that removes the familiar "anchor frame" constraint in favor of a generalized delayed decision analysis. This approach is referred to as adaptive frame selection using dynamic programming (AFS/DP). The AFS/DP algorithm determines optimal breakpoint locations using a perceptually-based performance measure, accommodates fixed bit-rate operation, and operates with fixed delay and low complexity. The AFS/DP algorithm has been implemented in the context of the MELP vocoder at 2400 bits/sec, and has demonstrated comparable performance to block VFR methods, with lower coder delay.

This paper describes our enhanced mixed excitation linear prediction (MELP) speech coder which is a candidate for the new U.S. Federal Standard at 2.4 kbits/s. The new coder is based on the MELP model, and it uses a number of enhancements as well as efficient quantization algorithms to improve performance while maintaining a low bit rate. In addition, the coder has been optimized for performance in acoustic background noise and in channel errors, as well as for efficient real-time implementation. Listening tests confirm that the enhanced 2.4 kbit/s MELP coder performs as well as the higher bit rate 4.8 kbit/s FS1016 CELP standard.

We propose three new adaptive noise suppression algorithms for enhancing noise-corrupted speech: smoothed spectral subtraction (SSS), vector quantization of line spectral frequencies (VQ-LSF), and modified Wiener filtering (MWF). SSS is an improved version of the well-known spectra subtraction algorithm, while the other two methods are based on generalised Wiener filtering. We have compared these three algorithms with each other and with spectral subtraction on both simulated noise and actual car noise. All three proposed methods perform substantially better than spectral subtraction, primarily because of the absence of any musical noise artifacts in the processed speech. Listening tests showed preference for MWF and SSS over VQ-LSF. Also, MWF provides a much higher mean opinion score (MOS) than does spectral subtraction. Finally, VQ-LSF provides a relatively good spectral match to the clean speech, and may, therefore, be better suited for speech recognition.

Traditional pitch-excited linear predictive coding (LPC) vocoders use a fully parametric model to efficiently encode the important information in human speech. These vocoders can produce intelligible speech at low data rates (800-2400 b/s), but they often sound synthetic and generate annoying artifacts such as buzzes, thumps, and tonal noises. These problems increase dramatically if acoustic background noise is present at the speech input. This paper presents a new mixed excitation LPC vocoder model that preserves the low bit rate of a fully parametric model but adds more free parameters to the excitation signal so that the synthesizer can mimic more characteristics of natural human speech. The new model also eliminates the traditional requirement for a binary voicing decision so that the vocoder performs well even in the presence of acoustic background noise. A 2400-b/s LPC vocoder based on this model has been developed and implemented in simulations and in a real-time system. Formal subjective testing of this coder confirms that it produces natural sounding speech even in a difficult noise environment. In fact, diagnostic acceptability measure (DAM) test scores show that the performance of the 2400-b/s mixed excitation LPC vocoder is close to that of the government standard 4800-b/s CELP coder.>

A number of improvements to the mixed excitation linear predictive coding (LPC) vocoder are presented. First, the authors have added more sophisticated frequency shaping of the pulse and noise in the mixture. They use a bandpass filter bank to attain a staircase approximation to any desired noise shape. Voicing strength in each frequency band is controlled by periodicity analysis of both the bandpass filtered speech and the bandpass speech envelope. Second, the authors have improved their pitch detection algorithm by using separate searches on the LPC residual and the input speech signal. Finally, they have added a fixed pulse shaping filter based on a spectrally flattened synthetic glottal pulse. The improved LPC vocoder performs much better in acoustic background noise, and it produces natural sounding speech in both quiet and noisy environments.>

The authors introduce a novel synthesizer structure for an LPC (linear predictive coding) vocoder which increases the clarity and naturalness of the output speech. This synthesizer enhances the usual excitations of either periodic pulses or white noise by allowing pulse/noise mixtures and aperiodic pulses, and thus can generate a wider range of possible speech signals. The control algorithms for this new model replace the traditional binary voicing decision with more robust periodicity, peakiness, and power level detectors, without a significant increase in bit rate. As a result, the vocoder produces synthetic speech which is free of the usual LPC synthesis artifacts, even at bit rates below 2400 bps.>

A new Linear Predictive Echo Canceller (LPEC) is presented which modifies the classical echo canceller structure to pre-whiten the input signal and hence have better convergence properties in the presence of highly correlated input signals such as speech. This structure is integrated with a 16 KBPS Residual Excited Linear Predictive (RELP) Vocoder [1] to allow significant execution speed savings by sharing common processing modules. The integrated LPEC has been implemented in the Vocoder Demonstration System (VDS-20) using a single TMS32020 processor and will be available at the conference for demonstrations.