This paper describes a 16-bit 250 MS/s ADC fabricated on 0.18 BiCMOS process. The has an integrated input buffer with new linearization technique that improves its distortion by 5-10 dB and lowers power consumption 70% relative to the state of art. It demonstrates background calibration correct residue amplifier (RA) gain errors lower consumption. summing node sampling (SNS) is based summing-node voltage using it corresponding estimate open loop gain. achieves SNDR 76.5 consumes 850 mW from...

We discuss a 14 bit 1 GS/s RF sampling pipelined ADC that utilizes correlation-based background calibration to correct the inter-stage gain, settling and memory errors. To improve linearity performance, employs input distortion cancellation digital technique compensate for non-linear charge injection (kick-back) from capacitors on driver. In addition, an effective dithering is embedded in signal break dependence of calibration's convergence amplitude. The fabricated 65 nm CMOS process has...

Marcos Zampieri, Shervin Malmasi, Nikola Ljubešić, Preslav Nakov, Ahmed Ali, Jörg Tiedemann, Yves Scherrer, Noëmi Aepli. Proceedings of the Fourth Workshop on NLP for Similar Languages, Varieties and Dialects (VarDial). 2017.

In this paper, we investigate different approaches for dialect identification in Arabic broadcast speech.These methods are based on phonetic and lexical features obtained from a speech recognition system, bottleneck using the i-vector framework.We studied both generative discriminative classifiers, combined these multi-class Support Vector Machine (SVM).We validated our results an Arabic/English language task, with accuracy of 100%.We also evaluated binary classifier to discriminate between...

In this paper we present a recipe and language resources for training testing Arabic speech recognition systems using the KALDI toolkit. We built prototype broadcast news system 200 hours GALE data that is publicly available through LDC. describe in detail decisions made building system: MADA toolkit text normalization vowelization; why use 36 phonemes; how generate pronunciations; build model. report results state-of-the-art modeling decoding techniques. The scripts are released on QCRI's...

We discuss a 12-b 18-GS/s analog-to-digital converter (ADC) implemented in 16-nm FinFET process. The ADC is composed of an integrated high-speed track-and-hold amplifier (THA) driving up to eight interleaved pipeline ADCs that employ open-loop inter-stage amplifiers. Up 10 GS/s, the THA operates at full sampling rate using non-interleaved single sample network, thereby eliminating interleaving time and bandwidth mismatch. Above programmed use two ping-ponged, or optional (2 + 1) randomized,...

This paper describes the Arabic MGB-3 Challenge - Speech Recognition in Wild. Unlike last year's MGB-2 Challenge, for which recognition task was based on more than 1,200 hours broadcast TV news recordings from Aljazeera programs, emphasises dialectal using a multi-genre collection of Egyptian YouTube videos. Seven genres were used data collection: comedy, cooking, family/kids, fashion, drama, sports, and science (TEDx). A total 16 videos, split evenly across different genres, divided into...

This paper describes a 14-bit, 125 MS/s IF/RF sampling pipelined A/D converter (ADC) that is implemented in 0.35 <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$muhbox m$</tex> BiCMOS process. The ADC has sample-and-hold circuit integrated the first pipeline stage, which removes need for dedicated amplifier (i.e., "SHA-less"). It also buffer turned off during hold clock phases to save power. To accurately estimate and minimize jitter, new...

Measuring the performance of automatic speech recognition (ASR) systems requires manually transcribed data in order to compute word error rate (WER), which is often time-consuming and expensive. In this paper, we propose a novel approach estimate WER, or e-WER, does not require gold-standard transcription test set. Our e-WER framework uses comprehensive set features: ASR recognised text, character results complement output, internal decoder features. We report for two features; black-box...

This paper describes the Arabic Multi-Genre Broadcast (MGB-2) Challenge for SLT-2016. Unlike last year's English MGB Challenge, which focused on recognition of diverse TV genres, this year, challenge has an emphasis handling diversity in dialect speech. Audio data comes from 19 distinct programmes Aljazeera channel between March 2005 and December 2015. Programmes are split into three groups: conversations, interviews, reports. A total 1,200 hours have been released with lightly supervised...

Dialect identification (DID) is a special case of general language (LID), but more challenging problem due to the linguistic similarity between dialects. In this paper, we propose an end-to-end DID system and Siamese neural network extract embeddings. We use both acoustic features for task on Arabic dialectal speech dataset: Multi-Genre Broadcast 3 (MGB-3). The was trained using three kinds features: Mel-Frequency Cepstral Coefficients (MFCCs), log Mel-scale Filter Bank energies (FBANK)...

Predicting the political bias and factuality of reporting entire news outlets are critical elements media profiling, which is an understudied but increasingly important research direction. The present level proliferation fake, biased, propagandistic content online has made it impossible to fact-check every single suspicious claim, either manually or automatically. Thus, been proposed profile look for those that likely publish fake biased content. This makes possible detect "fake news" moment...

We describe a 14-bit 2.5GS/s non-interleaved pipelined ADC that relies on correlation-based background calibrations to correct the inter-stage gain, settling (dynamic), kick-back and memory errors. A new technique is employed inject large dither signal input non-linear driver, another injected any residual non-linearity in pipeline. In order effect of aging comparators, calibration comparator offsets. The fabricated as dual 28nm CMOS process. An optional interleaved mode provided, where two...

In this paper, we describe Qatar Computing Research Institute's (QCRI) speech transcription system for the 2016 Dialectal Arabic Multi-Genre Broadcast (MGB-2) challenge. MGB-2 is a controlled evaluation using 1,200 hours audio with lightly supervised Our which was combination of three purely sequence trained recognition systems, achieved lowest WER 14.2% among nine participating teams. Key features our are: acoustic models recently introduced Lattice free Maximum Mutual Information (LF-MMI)...

Emotion recognition from speech signal based on deep learning is an active research area. Convolutional neural networks (CNNs) may be the dominant method in this In paper, we implement two architectures to address problem. The first architecture attention-based CNN-LSTM-DNN model. novel architecture, convo-lutional layers extract salient features and bi-directional long short-term memory (BLSTM) handle sequential phenomena of signal. This followed by attention layer, which extracts a summary...

This paper describes the fifth edition of Multi-Genre Broadcast Challenge (MGB-5), an evaluation focused on Arabic speech recognition and dialect identification. MGB-5 extends previous MGB-3 challenge in two ways: first it focuses Moroccan recognition; second granularity identification task is increased from 5 classes to 17, by collecting data 17 speaking countries. Both tasks use YouTube recordings provide a multi-genre multi-dialectal wild. transcription used about 13 hours transcribed...

Recent studies show that Gaussian mixture model (GMM) weights carry less, yet complimentary, information to GMM means for language and dialect recognition. However, state-of-the-art recognition systems usually do not use this information. In research, a non-negative factor analysis (NFA) approach is developed weight decomposition adaptation. This modeling, which conceptually simple computationally inexpensive, suggests new low-dimensional utterance representation method using similar of the...

High sample rate ADCs with high input bandwidth and low power consumption enable direct RF sampling, more integration, flexibility lower cost for communication, instrumentation other applications. The state of the art interleaved converters enables up to 10GS/s 12-14b resolution [1]-[4]. However, increase rate, number sub-ADCs tends increase, which degrades interleaving spurs due sampling time mismatch, increases capacitance, reduces bandwidth, ADC. Randomization helps alleviate impact...

In this paper, we describe a method to collect dialectal speech from YouTube videos create large-scale Dialect Identification (DID) dataset. Using method, collected Arabic known channels 17 speaking countries in the Middle East and Northern Africa. After refinement process, total of 3,000 hours was available for training DID systems, with an additional 57 development testing. For detailed evaluations, data divided into three sub-categories based on segment duration: short (less than 5s),...

End-to-end neural network systems for automatic speech recognition (ASR) are trained from acoustic features to text transcriptions.In contrast modular ASR systems, which contain separately-trained components modeling, pronunciation lexicon, and language the end-to-end paradigm is both conceptually simpler has potential benefit of training entire system on end task.However, such models more opaque: it not clear how interpret role different parts what information learns during training.In this...