Form and Function of Connectives in Chinese Conversational Speech

Automatic sentence segmentation of speech is important for enriching speech recognition output and aiding downstream language processing. This paper focuses on automatic sentence segmentation of speech in two different languages - English and Czech. For this task, we compare and combine three statistical models - HMM, maximum entropy, and a boosting-based model BoosTexter. All these approaches rely on both textual and prosodic information. We evaluate these methods on a corpus of multiparty meetings in English, and on a corpus of broadcast conversations in Czech, using both manual and speech recognition transcripts. The experiments show that superior results are achieved when all the three models are combined via posterior probability interpolation. We observe differences in terms of model performance between English and Czech, as well as the feature usage difference in prosodic models between the two languages. Overall, the analysis is important for porting sentence segmentation approaches from one language to another.

1. We used the pH-sensitive fluorescent dye 2',7'-bis(carboxyethyl)-5,6-carboxyfluorescein (BCECF) to study the regulation of intracellular pH (pHi) in single pyramidal neurons freshly isolated from the hippocampal CA1 region of immature (2- to 10-day-old) and more mature (21- to 30-day-old) rats. 2. Whether isolated from immature or mature rats, neurons had a broad range of initial pHi values (6.3-7.7) when the cells were examined in solutions buffered with Hepes and no CO2/HCO3-. The initial pHi distribution for neurons isolated from immature rats was best fitted with a Gaussian distribution with a mean of 6.95. In contrast, the initial pHi distribution for neurons isolated from mature rats was best fitted with the sum of two Gaussian distributions with means of 6.68 and 7.32. 3. When neurons with a relatively low initial pHi in Hepes-buffered solutions were acid loaded, pHi recovered very slowly. Neurons with a relatively high initial pHi recovered rapidly. The rate constant for the exponential pHi recovery increased with initial pHi. All pHi recoveries required Na+. 4. Both for neurons with a relatively high (> or = 7.05) and a relatively low (< 7.05) initial pHi, net acid extrusion rates (Jtotal = dpHi/dt x buffering power) decreased linearly with increasing pHi. Compared with the line for neurons with a relatively low initial pHi, that for neurons with a relatively high pHi had a significantly greater slope and was alkaline shifted by 0.6-0.7 pH units. 5. Removing external Na+ in the absence of CO2/HCO3- caused pHi to decrease by approximately 0.3 in neurons with a relatively low initial pHi, and by approximately 0.5 in neurons with a relatively high initial pHi. This initial acidification was followed by a slower, partial pHi recovery in approximately 32% of neurons with a relatively low initial pHi, but only approximately 14% of neurons with a relatively high pHi. 6. When exposed to CO2/HCO3-, all neurons initially acidified. Neurons with a relatively low initial pHi recovered to a pHi approximately 0.2 pH units greater than the initial value. Among neurons with higher initial pHi values, some did not recover at all, whereas others recovered to a value similar to or above the initial pHi. On average, the final CO2/HCO3- pHi for neurons with a relatively high initial pHi was similar to the pHi in Hepes buffer. Neurons with a relatively high pHi in Hepes buffer continued to be more alkaline (by approximately 0.2 pH units) in CO2/HCO3-. 7. When neurons with a relatively high initial pHi in Hepes (> or = 7.25) were exposed to CO2/HCO3- and then acid loaded, Jtotal values were more than twice the highest values observed in neurons with lower initial pHi values. Neurons with a moderate initial pHi in Hepes (7.05-7.24) had Jtotal values, at comparable pHi values, that were approximately 2-fold greater than for neurons with a relatively low initial pHi (< 7.05). 8. Thus, freshly isolated CA1 neurons of both mature and immature rats have a wide range of acid-base properties. Those with higher initial pHi values in a Hepes buffer tend to have greater Jtotal values in both Hepes and CO2/HCO3-, and tend to have higher steady-state pHi values in CO2/HCO3-.

Vowel reduction in conversational speech in French: The role of lexical factors

Voice quality has been defined variously in the literature ranging from states or postures of the glottis and vocal tract in general most broadly, to a narrower definition which refers to characteristics of vocal fold vibration during voiced phonation. Linguists have traditionally broken the voicing continuum into five basic categories based on roles they play in a language's phonology: (spread) voiceless, breathy, modal, creaky, (constricted) voiceless. Of these, the three central states, breathy, modal, creaky, are relevant to voice quality as discussed in this work. Voice qualities can be modelled as an interaction between subglottal pressure, degree of vocal fold approximation (aperture), longitudinal tension of the vocal folds (stiffness), and medial compression of the vocal folds (thickness). Breathy voicing is achieved with high glottal aperture, low stiffness, and low thickness, resulting in noise, low pitch, and increased spectral tilt. Creaky voicing has differing realizations depending on its linguistic role, prototypical creaky voice has low aperture, low stiffness, and high thickness, resulting in irregular and lower pitch, and decreased spectral tilt. Several other types of creaky voice quality exist including: 1) glottal fry, 2) multiply pulsed voice, and 3) nonconstricted creak. In this paper, we focus on creaky voice broadly-defined and concentrate on its distribution in North American English. While not contrastive, it plays an important role in phonology and a wide variety of other discourse, pragmatic, and social functions. In this context we present some of our current research into segmental and social factors relating to creaky voicing. We find a correlation between vowel height and creaky voicing. We also find evidence that voice quality is used by men to index gender in conversational speech. Our findings bear on the debate about the sociolinguistic uses of voice quality.

In adverse listening conditions, large and robust increases in intelligibility can be achieved by speaking clearly. The most striking differences between clear and conversational speech are associated with differences in speaking rate. To understand these differences, the intelligibility of speech in a variety of speaking modes was investigated at three different speaking rates. Talkers with significant speaking experience were asked to produce clear and conversational speech at slow, conversational, and quick rates. Previous studies show that the speaking rate for clear speech is roughly one-half that of conversational speech. Therefore, during training, the talkers were given feedback on their intelligibility in order to elicit the clearest possible speech at conversational and quick speaking rates. Talkers also recorded sentences in several other speaking modes such as soft, loud, and conversational with pauses inserted, as required for input to some automatic speech recognition systems. All speech materials recorded were nonsense sentences which provide no semantic context to aid listeners in identifying key words. The effects of changes in speaking rate on intelligibility of various speaking modes for normal hearing listeners in a background of wideband noise are discussed. [Work supported by NIH.]

A significant source of errors in Automatic Speech Recognition (ASR) systems is due to pronunciation variations which occur in spontaneous and conversational speech. Usually ASR systems use a finite lexicon that provides one or more pronunciations for each word. In this paper, we focus on learning a similarity function between two pronunciations. The pronunciations can be the canonical and the surface pronunciations of the same word or they can be two surface pronunciations of different words. This task generalizes problems such as lexical access (the problem of learning the mapping between words and their possible pronunciations), and defining word neighborhoods. It can also be used to dynamically increase the size of the pronunciation lexicon, or in predicting ASR errors. We propose two methods, which are based on recurrent neural networks, to learn the similarity function. The first is based on binary classification, and the second is based on learning the ranking of the pronunciations. We demonstrate the efficiency of our approach on the task of lexical access using a subset of the Switchboard conversational speech corpus. Results suggest that on this task our methods are superior to previous methods which are based on graphical Bayesian methods.

Wireless communication has become increasingly popular in recent years due to its mobility, portability and efficient service. Quality Voice Communication is the basic need of all such devices. Due to the significant spectrum demand, the bandwidth allocated to Professional Radio devices is decreasing. This results in lowering voice/ speech quality. Due to limited bandwidth, the design of Radio is changing. The main change is shifting from Analogue signal processing to Digital signal processing of speech and radio signals. The conversion of analogue-natured speech to digital form itself creates loss and faithfulness. Such Radio includes a voice coder for analogue to digital conversion of speech. Human spoken language, speaker (male or female), and pronunciation play an essential role in direct voice and Radio communication. It is the primary object which decides the design of low bandwidth coder. This paper deals with Marathi language characteristics and the interdependence of language constructs that affects voice/speech quality. In this paper, we extracted features of the Marathi language using PRAAT software and analyzed its constructs using statistical tools. The actual testing on Radio is carried out. The speech quality observed during such testing is correlated with statistical results. It is shown that the standard deviation in pitch and formants in male speech is less (average 35%) compared to the standard deviation in pitch and formants in female speech. This results in lowering the quality of male speech compared to female speech. It is also shown that some vowels and combinations of certain vowels and consonants produce poor speech quality compared to others.

Ericksonian hypnotic procedures can be used by a wide range of alternative practitioners. They do not require the traditional trance and can be adapted based on the client's health needs, using conversational speech. Whether practitioners are aware of it or not, they give suggestions. This article presents ways alternative practitioners can readily use to make sure the suggestions they use are positive and support healing.Health care practitioners can easily learn to recognize the waking trance experience and begin to employ the use of therapeutic suggestion in their usual interactions with patients. This article presents a brief introduction to the therapeutic use of hypnosis and direct and indirect suggestions. They are tools to be utilized within the auspices of one's own profession. To become proficient in utilizing these approaches, formal training with reputable organizations is necessary.

The purpose of this study was to identify the manifestation of metaphorical asymmetries in the expression of joy and sadness across English, Kazakh, and Chinese and how these asymmetries reflect cultural, cognitive, and sociolinguistic frameworks. Emotions like joy and sadness are frequently challenging to convey explicitly, and metaphors function as vital instruments for expressing these intricate experiences. This study delineates universal and culture-specific metaphorical constructs, illustrating that joy and sadness often utilise identical source domains (light / darkness, rise / fall, burden / release), yet manifest asymmetrically regarding intensity, prominence, and symbolic elaboration. The research utilises cognitive linguistics, psycholinguistics methodology, and conceptual metaphor theory to analyse a multilingual corpus comprising literary texts, journalistic discourse, conversational speech, and digital communication. A statistical analysis of over 1,000 metaphorical expressions revealed that spatial metaphors were the most prevalent across all languages (30.2%), succeeded by temporal metaphors (20.1%) and emotion-as- sensation metaphors (16.7%). Significant cross-linguistic variation was identified. English data indicated a pronounced inclination towards metaphors of brightness and verticality (p=0.021), Kazakh texts revealed a statistically significant prevalence of burden metaphors (U=112.5, p=0.008), and Chinese metaphors for sadness demonstrated strong correlations with weather and imbalance imagery. Correlation analysis indicated that in Kazakh, the metaphor “sadness is a burden” frequently co-occurred with intensifiers (r=0.61, p=0.013), illustrating that metaphors are pragmatically reinforced in discourse. These findings validate the claim that metaphors convey emotions and represent socio-cultural perspectives. Kazakh metaphors illustrate a profound relationship with nature and kinship, and Chinese metaphors highlight harmony and balance, whereas English metaphors prioritise individuality and psychological conditions. The research shows that metaphorical asymmetry serves as an effective framework for comprehending the conceptualisation of emotions across languages. In addition to cultural influence, metaphors serve as semiotic evidence that facilitates the reconstruction of cultural orientations and cognitive preferences. These findings are pertinent to cognitive linguistics, translation studies, and intercultural communication, underscoring the necessity of including metaphorical asymmetry into language instruction and cross- cultural study.

The article deals with the analysis of semantic compression application in popular English talk shows. The purpose of this paper is to define language compression, to outline the expediency of its use in English media discourse, to name ways of compression implementation in popular talk shows and to consider the level of influence of principle of linguistic efforts economy on achieving communicative goals between interlocutors. Basic information on the concept of linguistic compression and features of its use at different language levels, in particular phonetic, morphological, lexical, semantic and syntactic is presented. Linguistic compression is used in oral and written discourse, in formal and informal texts. The text gives valuable information on the structure and characteristics of media discourse, its grammatical, lexical and phonetic features. The article focuses on a detailed analysis of the main linguistic means that serve to reduce the symbolic structure of language, which leads to the brevity of expression, contains the data on key properties of linguistic compression that help make the text more concise. Special attention is paid to the phonetic methods of linguistic compression in English talk shows: it is manifested through such means as reduction, assimilation, loss, fusion and other processes. Among the syntactic means of compression, such methods used for linguistic efforts economy as contamination, segmentation and parceling, ellipse, univerbation, abbreviation, composite words, inclusion and use of foreign words or components are analyzed. The article points out the importance of adequate transfer of information with preservation of expressive nuances and observance of stylistic language norms. The article is of interest for further researches on the topic of linguistic efforts economy in oral and written speech styles, can be used in researches in linguistics, lexicology, phonetics.

Change in singing voice production, objectively measured

日本語話者が発した音声を, 無響室内で1m前方と外耳道の入口部近傍の等価音圧レベルと等価騒音レベルを測定し, これらを比較検討した。対象は, 男性10名と女性10名の計20名とした。その結果, 日本語音声の1m前方の等価音圧レベルは約60dBで, 等価騒音レベルは約56dBとなった。それらの差は約4dBであった。外耳道の入口部近傍の耳元に到達する音声の等価音圧レベルは約74dB, 等価騒音レベルは約69dBとなることが示された。これらの結果は, 音場で補聴器の適合状態を評価したりするときの条件設定や, 補聴器などによる聴力障害との関係を検討するための基礎資料となると思われた。

We present an overview of the data collection and transcription efforts for the COnversational Speech In Noisy Environments (COSINE) corpus. The corpus is a set of multi-party conversations recorded in real world environments with background noise that can be used to train noise-robust speech recognition systems. We explain the motivation for creating such a corpus and describe the resulting audio recordings and transcriptions that comprise the corpus. These recordings include a 4-channel array and close-talking, far-field, and throat microphones on separate synchronized channels, allowing for unique algorithm research.

In this work, a novel training scheme for generating bottleneck features from deep neural networks is proposed. A stack of denoising auto-encoders is first trained in a layer-wise, unsupervised manner. Afterwards, the bottleneck layer and an additional layer are added and the whole network is fine-tuned to predict target phoneme states. We perform experiments on a Cantonese conversational telephone speech corpus and find that increasing the number of auto-encoders in the network produces more useful features, but requires pre-training, especially when little training data is available. Using more unlabeled data for pre-training only yields additional gains. Evaluations on larger datasets and on different system setups demonstrate the general applicability of our approach. In terms of word error rate, relative improvements of 9.2% (Cantonese, ML training), 9.3% (Tagalog, BMMI-SAT training), 12% (Tagalog, confusion network combinations with MFCCs), and 8.7% (Switchboard) are achieved.

We have recently proposed a new acoustic model based on probabilistic linear discriminant analysis (PLDA) which enjoys the flexibility of using higher dimensional acoustic features, and is more capable to capture the intra-frame feature correlations. In this paper, we investigate the use of bottleneck features obtained from a deep neural network (DNN) for the PLDA-based acoustic model. Experiments were performed on the Switchboard dataset — a large vocabulary conversational telephone speech corpus. We observe significant word error reduction by using the bottleneck features. In addition, we have also compared the PLDA-based acoustic model to three others using Gaussian mixture models (GMMs), subspace GMMs and hybrid deep neural networks (DNNs), and PLDA can achieve comparable or slightly higher recognition accuracy from our experiments.