Professional Documents
Culture Documents
Human Language Technologies For Under Resourced African Languages Design Challenges and Prospects 1st Edition Moses Effiong Ekpenyong (Eds.)
Human Language Technologies For Under Resourced African Languages Design Challenges and Prospects 1st Edition Moses Effiong Ekpenyong (Eds.)
https://textbookfull.com/product/using-comparable-corpora-for-
under-resourced-areas-of-machine-translation-inguna-skadina/
https://textbookfull.com/product/human-language-technology-
challenges-for-computer-science-and-linguistics-zygmunt-vetulani/
https://textbookfull.com/product/biofuels-technology-challenges-
and-prospects-1st-edition-avinash-kumar-agarwal/
https://textbookfull.com/product/combustion-for-power-generation-
and-transportation-technology-challenges-and-prospects-1st-
edition-avinash-kumar-agarwal/
Agroforestry to Combat Global Challenges: Current
Prospects and Future Challenges 1st Edition Hanuman
Singh Jatav
https://textbookfull.com/product/agroforestry-to-combat-global-
challenges-current-prospects-and-future-challenges-1st-edition-
hanuman-singh-jatav/
https://textbookfull.com/product/human-vaccines-emerging-
technologies-in-design-and-development-1st-edition-kayvon-
modjarrad/
https://textbookfull.com/product/challenges-of-globalization-and-
prospects-for-an-inter-civilizational-world-order-ino-rossi/
https://textbookfull.com/product/green-public-procurement-under-
wto-law-experience-of-the-eu-and-prospects-for-switzerland-rika-
koch/
https://textbookfull.com/product/biomaterials-for-organ-and-
tissue-regeneration-new-technologies-and-future-prospects-1st-
edition-nihal-vrana-editor/
SPRINGER BRIEFS IN ELEC TRIC AL AND
COMPUTER ENGINEERING SPEECH TECHNOLOGY
Human Language
Technologies for
Under-Resourced
African Languages
Design, Challenges,
and Prospects
123
SpringerBriefs in Electrical and Computer
Engineering
Speech Technology
Series editor
Amy Neustein, Fort Lee, NJ, USA
Editor’s Note
The authors of this series have been hand-selected. They comprise some of the most
outstanding scientists—drawn from academia and private industry—whose
research is marked by its novelty, applicability, and practicality in providing broad
based speech solutions. The SpringerBriefs in Speech Technology series provides
the latest findings in speech technology gleaned from comprehensive literature
reviews and empirical investigations that are performed in both laboratory and real
life settings. Some of the topics covered in this series include the presentation of
real life commercial deployment of spoken dialog systems, contemporary methods
of speech parameterization, developments in information security for automated
speech, forensic speaker recognition, use of sophisticated speech analytics in call
centers, and an exploration of new methods of soft computing for improving
human-computer interaction. Those in academia, the private sector, the self service
industry, law enforcement, and government intelligence, are among the principal
audience for this series, which is designed to serve as an important and essential
reference guide for speech developers, system designers, speech engineers, linguists
and others. In particular, a major audience of readers will consist of researchers and
technical experts in the automated call center industry where speech processing is a
key component to the functioning of customer care contact centers.
Human Language
Technologies for
Under-Resourced African
Languages
Design, Challenges, and Prospects
Editor
Moses Effiong Ekpenyong
Department of Computer Science
University of Uyo
Uyo, Akwa Ibom State, Nigeria
Speech technology has become almost taken for granted in everyday life.
Fundamental techniques of automatic speech recognition (ASR), text to speech syn-
thesis (TTS), and speaker recognition and verification have become available as
components of commercial interactive search and consumer service agents, which
are implemented as smart loudspeakers and supported by complex databases pro-
cessed with machine learning techniques. Older application fields of public
announcements in travel hubs and dictation software and reading applications for
the visually impaired are becoming more widespread. Increasingly, speech is under-
stood not only as spoken language but as a multimodal complex of parallel synchro-
nised data streams of audio-visual information from spoken language itself together
with body movement: facial mimicry, manual gesture, and posture.
To a large extent, the world of speech technology applications is still focused on
the languages of regions with major research and development resources, such as
the European Union, North America, India, China, and Japan, but the proceedings
of international conference series such as the Language Resources and Evaluation
Conferences (LREC), InterSpeech, the IEEE International Conference on Acoustics,
Speech and Signal Processing (ICASSP), the Language and Technology Conference
(LTC), and Oriental COCOSDA have shown a rapid increase in papers on the less
resourced languages of Asia, Africa, and the Americas. Even so, these developments
are largely driven by technologies developed in the wealthier nations on the basis of
their standard, ‘commercially interesting’ languages.
A central part of the speech system development process is the creation of lan-
guage and speech resources for the languages concerned. This central part is to a
large extent also the most complex part. First, the process splits into many different
subprocesses: the pre-recording process of data resource design, the scenario-spe-
cific recording processes, and the post-recording procedures of annotation, analysis
with machine learning procedures, evaluation, archiving, and dissemination.
Second, for each of the subprocesses, the appropriate tool resources need to be
developed. A key point to remember is that data type and quality are essentially
functions of the tools used in the acquisition and development process, just as
v
vi Foreword
The authors of this volume have set themselves the goal of fulfilling the above
criteria in order to face the challenge of integrating local languages, in this case in
the region of South East Nigeria, into the digital community. The individual chap-
ters address issues in system development and its relation to the digital economy,
and to digital services for health and education. These commendable endeavours
are already showing fruit and will certainly develop into valuable contributions to
overcoming the digital divide. The contributions will provide effective data and
tool resource models, not only for Nigeria but also for areas where other Niger-
Congo languages are spoken. On this basis, integration of West African speech
technology development into the international speech technology community is
now well on its way.
ix
x Preface
Index������������������������������������������������������������������������������������������������������������������ 127
xi
About the Author
xiii
Chapter 1
Adaptive Template-Based Front End for Tone
Language Speech Synthesis
1.1 Introduction
Text-to-speech (TTS) synthesis has transformed dramatically over the last couple of
decades, such that most current TTS systems relatively apply data-driven tech-
niques, instead of rule-based techniques – one of the earliest language-specific tech-
niques that suffered the challenges of synthesizing unrestricted domain texts. A TTS
system is basically composed of two parts: the front end or high-level synthesis
(Natural Language Processing (NLP) phase) and the back end or low level synthesis
(Digital Signal Processing (DSP) phase). The front end is responsible for gathering
input in various forms from the user, and processing it to conform to a specification
the back end can use. It is often regarded as the interface between the user and the
back end. The back end, also referred to as the waveform generator, is responsible
for the conversion of the linguistic representation into sounds. In some systems, this
part consists of the computation of the target prosody (pitch contour, phoneme dura-
tions), which is then imposed on the output speech. Figure 1.1 summarizes the func-
tional blocks of these two phases.
The application of Hidden Markov Models (HMMs) in speech synthesis is still
gaining prominence. In conventional statistical parametric speech synthesis, distinct
HMM for each context combination is typically employed to represent probability
densities of speech parameters of input texts, from which speech parameters are
generated to maximize the output probabilities, with a final reconstruction of the
speech waveform. Parametric speech synthesis techniques rely on full context
acoustic models generated by language front ends (Aylett et al. 2014), and respon-
sible for analyzing the linguistic and phonetic structure of the language.
Text Prosody
Text Text pre- and pronuncia- NLP phase
Annotations
processing tion prediction
Prosodic labels
Phone labels
Waveform DSP phase
synthesis
Speech
Most current voice services employ concatenation techniques with units at least the
size of words or phrases. These techniques which provide near natural speech qual-
ity are limited by low flexibility because of their dependence on pre-recorded speech
1 Adaptive Template-Based Front End for Tone Language Speech Synthesis 3
method may be usable to some extent when languages are prosodically and phone-
mically the same, but severe problems arise when languages are typologically very
dissimilar. For example, intonation languages, for which TTS systems have been
typically developed, pose very different problems from tone languages, e.g., Ibibio.
Even systems for Chinese, also a tone language, cannot be generalised in this way
because the East Asian languages in general have phonemic tone, whereas African
languages have a broad spectrum of morphological tone functionalities in addition
to phonemic tone. For instance, in Ibibio, the subcategory of proximal/distal (tem-
porally near or far) tense is marked by Low-High/High-Low (LH/HL) tones on the
tense morphemes (e.g. used in the context (proximal/distal
tense) ‘I (will go)/went to the market tomorrow/yesterday’.
Pitch therefore has a hard mandatory semantic function in tone languages rather
than a soft pragmatic function in intonation languages (Gibbon et al. 2006;
Ekpenyong and Udoh 2014). Hence, if tone were orthographically marked, tone-
morpheme combinations could be used to capture this problem in unit selection and
in text pre-processing, with a resulting explosion of the corpus size. However, the
problem is compounded for Ibibio by the lack of orthographic tone marking, mak-
ing morphological tone assignment effectively an Artificial Intelligence (AI) com-
plete problem (Gibbon et al. 2006; Ekpenyong et al. 2014), and requiring extensive
background world knowledge, i.e., heuristic guessing algorithms for morphological
tone assignment. Again, the positional dependence of tone values on the terraced
tone patterning generated by automatic and non-automatic downstep in many
African languages determines a further combinatorial explosion of pitch patterning
(Gibbon 2001). In a number of African languages, the number of inflected word
forms is far larger than for languages like English or Chinese due to the agglutina-
tive inflectional morphology and complex subject-verb-object person concord. This
adds further complexity to morphological tone assignment and produces problems
of text corpus sparseness in the former, a specific case of the sparse data problem,
prevalent in corpus-based approaches (Saruladha et al. 2010).
(i) difficulty in porting the system across language domains and platforms;
(ii) acquisition of the knowledge base or development of the language dependent
representations is highly labour-intensive;
(iii) linguistic models are not so interesting for language and speech processing,
which limits their reproduction;
(iv) there are no underlying and unifying linguistic theories to sufficiently and pre-
cisely represent all levels of linguistic phenomena – phonetic, morpho-
syntactic, syntactic, discourse, etc.
One classical example of RB systems is the domain-specific expert system that
utilizes rules to make deductions or decisions. Here knowledge is represented as a
set of rules and data as a set of facts. A rule engine is then exploited to compare each
rule (in the knowledge base) with the facts.
Data-driven (empirical) System Automatically derive or learn the linguistic units
from exemplars, thus precluding the acquisition of prior linguistic knowledge.
Today, there is huge availability of speech and language data (audio, speech, video
and text) for well-resourced languages on the internet. These data can be inexpen-
sively harvested to rapidly create data-driven systems that are portable across lan-
guage domains and platforms. DD methods require a speech database, ideally with
labels or can be generated automatically, if not available. DD systems are more
desirable because they require less human expert knowledge and are flexible to
adapt when deployed. They are therefore attractive because speech and language
processing problems have evolved efficient solutions to the storage, indexing,
retrieval and processing of available data. As the number of available data increase,
more representative exemplars are generated, and more accurate results are pro-
duced. The role of a DD system would now be to efficiently store, index, retrieve
and process the relevant units of speech or text. A typical example of a DD system
is a translation system with access to huge amount of translations between two lan-
guages. The role of the DD system would involve matching the input (source lan-
guage) with the translations in its database, retrieving appropriate units in the target
language and possibly combining them to generate a translation. As more transla-
tions become available, the degree of ambiguity and data sparseness decreases, and
more accurate translations and equivalents are produced. Hence, the resulting trans-
lation becomes more precise. In constructing a DD system, the use of HMMs has
largely arisen (Tokuda et al. 2000). HMMs have successfully been applied to mod-
eling the sequence of speech spectra in speech processing systems, but much of the
advances in speech synthesis have been borrowed from the field of speech recogni-
tion. The field of machine learning has increasingly exploited data-driven approaches
where large databases act as implicit knowledge sources, rather than explicit rules
manually written by experts. Machine learning techniques are preferred in situa-
tions where engineering approaches such as hand-crafted models cannot cope with
the problem complexity, and are usually classified into three broad categories:
Predictive or supervised here the goal is to learn a generic rule that maps inputs, i,
to outputs, o, given a labeled set of input-output pairs, T = {( ik , ok )}k =1
N
6 M.E. Ekpenyong
T = {( ik , ok )}k =1 , where T is the training set, and N is the number of training exem-
N
plars. A supervised learning algorithm then analyzes the training data and produces
an inferred function, which can be used to map new examples. An optimal scenario
will permit the algorithm to correctly detect unseen labels or instances, and this
requires the learning algorithm to generalize from the training data to unseen cases
in a “reasonable” way.
Descriptive or unsupervised unsupervised learning can be a goal in itself. In this
case, only the inputs, T = {ik }k =1 , T = {ik }k =1 are given, and the goal is to discover
N N
T1: Language-specific
generate
couple
Ibibio is a Lower Cross tone language, from the New Benue Congo language fam-
ily, spoken in the Southeast Coastal region of Nigeria, which native speakers are
predominantly found in Akwa Ibom State. Ibibio represents the fourth largest speak-
ing group in Nigeria, and consists of about four million speakers. The speech char-
acteristics of Ibibio concern the rules governing the production of sounds in the
language, and are discussed in the following subsections.
Vowel and consonant system The Ibibio vowel and consonant systems are pre-
sented in Tables 1.1 and 1.2, respectively.
The long and short vowels are similar in their tone levels, but different in the
sense that the long vowels are merely lengthened versions of their short counterpart,
and they do not bear the rising and falling tones.
Syllable Structure One of the most important aspects of phonology is the structural
representation of sound patterns above the levels of phonemes (Ekpenyong and
Udoh 2014). The Ibibio speech system is built around its syllable structure consist-
ing of a single onset-vowel or nasal prefix and a rhyme consonant (Urua 2000). The
syllable structure of Ibibio is distributed as follows (Ekpenyong and Udoh 2014):
(i) n̄(N) syllable structure: This is a syllabic nasal, and is homoganic with conso-
nants such as ‘wing’, ‘book’, ‘society’;
(ii) V syllable structure: This can be observed where a vowel occurs as a prefix in
a word as in ‘naval’;
(iii) CV syllable structure: This can be seen in words like ‘stand’, ‘come’;
(iv) CVC syllable structure: This structure can be seen in words like ‘throw’,
‘cook’.
(v) CGV syllable structure: This is a consonant-glide vowel structure such as
‘sit’.
(vi) CGVC syllable structure: This can be observed in words like
‘split’.
(vii) C V syllable structure: This is seen in the only consonant structure cluster in
Ibibio as in ‘play’, ‘stop’.
1 Adaptive Template-Based Front End for Tone Language Speech Synthesis 9
Initial processing of the input (text) utterance for an Ibibio TTS front end was
accomplished using a text-to-segments framework and implemented with Speect – a
multilingual speech synthesis system (Louw 2008). Speect offers application pro-
gramming interfaces, as well as environment for research and development of TTS
systems and voices. We obtained from this implementation, heterogeneous relation
graphs (HRGs) describing the linguistic structures of the language. HRG was devel-
oped for use in speech synthesis systems, e.g., in Festival, (Black, Taylor and Caley,
1996–1999), and its design addresses the specific needs of such a system. HRG is
useful because of the specific formalism it provides for preserving linguistic infor-
mation in speech synthesis systems. This formalism differentiates itself from other
formalisms used in speech and language processing. The linguistic data processed
in a synthesis system is linguistically heterogeneous, so, rather than dealing with
syntax or phonology independently, synthesizers can be involved in text analysis,
syntactic analysis, morphology, phonology, phonetics, prosody, articulatory control
and acoustics. Hence, to encourage a rich and robust structure that would be useful
not only for text/speech processing, it is essential for a synthesis system to store
representations of these different types of linguistic information in a single
formalism.
Input to the Speect system were recorded utterance and phoneme inventory (see
T1, in Fig. 1.2). The recorded utterance have been used in larger Ibibio synthesis
experiments (Ekpenyong 2013) and were processed from resources collected in
three different projects namely, the West African Language Archive (WALA) proj-
ect, the Local Language Speech Technology Initiative (LLSTI, http://www.llsti.org)
project, and a World Bank/Science and Technology Education Post-Basic (STEP-B)
project. The processed corpus contained a total of 1140 utterances, samples of
which are shown in Fig. 1.3.
To ensure efficient speech preprocessing, and avoid loss of information, the
speech corpus was coded using the Speech Assessment Method for Phonetic
Alphabet (SAMPA) notations, and tailored to suit the ergonomic needs of the lan-
1. bONakam kuukpamba
2. akefeefeRe ajak ikOt abasi
3. eJe amaanam aNwaNa ke mme owo enie ntreubOk ke usVN OmmO keed keed
4. abasi amaasiak usVN ubOkkO OnO nditO isred
5. ejIn OmO adiben eJe akaaisaN
6. eteidVN kiNsidi
7. idooRo akpanikO owo edinIm mbo ke owo emi ataaRa Nwet abasi ke ido nsu
isibOOhO ufen
:::
1137. daNa ebod odo akekan odu OJVN asaNa OJON akewOd ke ekON ikikemme usen
1138. imam isinemme akaN iba mme owo emaesak imam tutu eJe ebeek OmmO NkaN
1139. ete ifOn akedo amaasak mmOONOjId afeRe ke mfVk nte ndiON edIm
1140. daNa ekedikOppO mbVk ebod imam amamaana asakka
The context-dependent features for English (Zen 2006) were modified for tone lan-
guages. The modified version is shown in Fig. 1.6. Features related to stress lan-
guages were suppressed using a ‘not applicable’ or NULL letter ‘x’.
In Fig. 1.6, p1, p2, …, t3, represent linguistic features or model states of the text
utterances (recorded corpus data), which models features such as the phonetic con-
text, syllable, word, phrase and utterance statistics, prosodic and tone patterning of
the intended language. Table 1.4 describes the linguistic features of the HMM labels
in Fig. 1.6.
12 M.E. Ekpenyong
{duration} p1^ p2 − p3 + p4 = p5 @ p6 _ p7
/ A : x _ x _ a3 / B : x − x − b3 @b4 − b5 &b6 − b7 # x − x$x − x! x − x; x − x | b16/ C : c1 + c2 + c3
/ D : d1 _ d 2 / E : e1 + e2 @ e3 + e4 & e5 + e6 # e7 + e8 / F : f1 _ f 2
/ G : g1 _ g 2 / H : h1 = h2 @ h3 = h4 | x / I : i1 _ i2
/ J : j1 + j2 − j3 / TL: t1 _ TC: t 2 _ TR: t 3
To obtain statistics and position related information for other linguistic features,
excluding tone features, a Shell script was then written to flatten the HRG files into
a single text file. A typical output of processed HRGs revealing only the syllable
annotations of sample utterances in Fig. 1.3 is given in Fig. 1.7.
Notice in Fig. 1.6, that symbols are used as boundary separators. In our case, ‘#’
represents word boundary separator, ‘-’ represents syllable boundary separator, and
‘|’ represents phrase boundary separator.
To obtain statistics for tone related features, close copy annotations of the input
utterances (see Fig. 1.3) was manually carried out for tones. The annotated v ersion/
file was then automatically processed into a syllabified equivalent of Fig. 1.7. The
output of this processing is shown in Fig. 1.8.
Notice in Fig. 1.8 that only tones are processed. In the case of Ibibio, numbers
(1–5) were used to represent the various tones as follows: 1 for High (H) tone, 2 for
Low (L) tone, 3 for Down-stepped (!) tone, 4 for High Low (HL) or Falling tone,
and, 5 for Low High (LH) or Rising tone. Zeros (0’s) were automatically generated
to fill the consonants slots.
To implement context-dependent tone modelling, we propose a data-driven tech-
nique to features specification to make our framework flexible at the front end – and
minimise the hard-coding of linguistic evidence into the synthesiser, hence, making
the system generic and adaptable to other tone languages. Figure 1.9 shows the
proposed context-dependent HMM algorithm for tone and prosody features label-
ling, and is expected to be fully implemented in future research.
From Fig. 1.9, we can conveniently extract the tone and prosodic features of any
corpus by clustering the input patterns and emitting the expected transitions at each
state of the HMM, thus:
r r r r
Tlabel = θ 0f,tone (i ,1) + θ tone (i ,1) +…+ θ tone(i ,n −1) + θ cf(i ,n −1),tonepat (i , n −1) +
r r r r s
θ tone (i ,n ) + θ cf(i ,n ),tonepat (i ,n ) + θ tone(i ,n +1) +…+ θ tone (i , N ) + θCb+1,tone(i , N ) +
s s s s s (1.1)
θ tpros (i , N ) +…+ θ pros (i ,n +1) + θ cb(i ,n ),tonepat (i ,n ) + θ pros (i ,n ) + θ cb(i ,n −1),tonepat (i ,n −1) +
s s
θ pros (i ,n −1) +…+ θ pros (i ,1)
Equation (1.1) is useful for modelling the state features of a HMM-based tone
language synthesis system. It describes the context-dependent features and details
the prosodic factors necessary for tone language synthesis. The prosodic features
1 Adaptive Template-Based Front End for Tone Language Speech Synthesis 15
1. #bO-Na-kam|kuu -kpam-ba
2. #a-ke-fee-fe-Re#a-jak#i-kOt#a -ba-si
3. #e-Je#a-maa-nam#aN -wa-Na#ke#m -me#o-wo#e-nie#n-treu-bOk#ke#u -sVN#Om-
mO#keed#keed
4. #a-ba-si#a-maa-siak#u -sVN#u -bOk-kO#O-nO#n-di-tO#i-sred
5. #e-jIn#O-mO#a -di-ben#e-Je#a-kaai-saN
6. #e-tei-dVN#kiN -si-di
7. #i-doo-Ro#a-kpa-ni-kO#o-wo#e-di-nIm#m -bo#ke#o -wo#e-mi#a-taa-Ra#N-wet#a -
ba-si#ke#i -do#n-su#i-si-bOO-hO#u-fen
:::
1137. #da-Na#e-bod#o-do#a-ke-kan#o-du#O-JVN#a -sa-Na#O-JON#a -ke-wOd#ke#e -
kON#i -ki-kem-me#u-sen
1138. #i-mam#i -si-nem-me#a-kaN#i-ba#m-me#o-wo#e-mae-sak#i-mam#tu -tu#e-Je#e-
beek#Om -mO#N -kaN
1139. #e-te#i-fOn#a -ke-do#a-ma-a-sak#m-mOO-NO-jId#a-fe-Re#ke#m -fVk#n -te#n-
diON#e -dIm
1140. #da-Na#e-ke-di-kOp-pO#m-bVk#e-bod#i -mam#a -ma-maa-na#a-sak-ka
Fig. 1.10 Context dependent labels for the input utterance ‘ete idVN kiNsidi’. (a). Phonetic con-
text model; (b). Full context model for first three phoneme-sequences
describing each vector of Eq. (1.1) are given in Table 1.4. A Python script was writ-
ten to extract the features in Table 1.4, from a text corpus, and a snippet of the full
context model for a sample sentence is given in Fig. 1.10b
To generate the HTS label (or lab) files, context interactions between the various
features in Fig. 1.7 were modeled in a supervised manner using a simple FST. First,
the speech duration and phonetic contexts to establish the initial quin-phone labels
was dealt with. In labeling the speech durations, we do not manually annotate the
speech corpus, but assume that each phone in a given utterance has same duration.
This initial assumption was used to pre-annotate the speech corpus and was achieved
by dividing the total utterance’s duration by the total number of phonemes in that
utterance. This gave all phone durations equal lengths, with fixed incremental fac-
tors (Fig. 1.10a). We have shown that the resultant voice from quin-phone labels
sounds poor, but improved greatly when a rich/full context model was used
(Ekpenyong et al. 2014). A full context-dependent model for the first two phoneme
sequences in Fig. 1.10a is shown in Fig. 1.10b. To obtain consistent duration model-
ing, a re-alignment process using the Viterbi algorithm was adopted. Viterbi algo-
rithm is the most common algorithm for implementing n-gram search. It is an
efficient dynamic search technique that avoids the polynomial expansion of a
breath-first search, by ‘trimming’ the search tree at each level using the best n
MLEs. The re-alignment process was achieved using HMM edit (HLEd) tool of the
HTS toolkit.
Another random document with
no related content on Scribd:
Lameyran
Gmelin,
and Fremy,
Gas. Pflüger. from Lassaigne. Vogel. Raiset.
from
clover
clover.
CO 60–80 5 29 5 27 74.23
CO 28–40
CH 15 6 15 48 23.46
HS 80 80
O 14.7
N 50.3 25 2.22
The most elaborate observations on this subject are those made by
Lungwitz on the different aliments kept in closed vessels at the body
temperature, and on similar agents fed for days as an exclusive
aliment to oxen provided with a fistula of the rumen for purposes of
collection. He found carbon dioxide to be the predominating gas in
all cases, but that it was especially so in extreme tympanies and
varied much with the nature of the food. The following table gives
results:
Percentage of CO2.
Buckwheat (Polygonum fagopyrum) 80
Alfalfa (Mendicago Sativa) 70–80
Clover (Trifolium pratense) 70–80
Meadow grass 70–80
Indian corn (Zea Maïs) 70–80
Spurry (Spergula arvensis) 70–80
Hay of alfalfa or clover 70–80
Oats with cut straw 70–80
Yellow Lupin (Lupinus luteus) 60–70
Vetch (Vicia sativa) 60–70
Oats cut green 60–70
Potato tops 60–70
Potatoes 60–70
Meadow hay 60–70
Leaves of beet 50–60
Leaves of radish 50–60
Cabbage 40–50
The marsh gas varied from 16 to 39 per cent., being especially
abundant in cases of abstinence. It should, therefore, be in large
amount in the tympanies which accompany febrile and other chronic
affections. Hydrogen sulphide was found only in traces, recognizable
by blackening paper saturated with acetate of lead. Oxygen and
nitrogen were in small amount and were attributed to air swallowed
with the food. In the work of fermentation the oxygen may be
entirely used up.
Lesions. These are in the main the result of compression of the
different organs, by the overdistended rumen. Rupture of the rumen
is frequent. The abdominal organs are generally bloodless, the liver
and spleen shrunken and pale, though sometimes the seat of
congestion or even hemorrhage. Ecchymoses are common on the
peritoneum. The right heart and lungs are gorged with black blood,
clotted loosely, and reddening on exposure. The right auricle has
been found ruptured. Pleura, pericardium and endocardium are
ecchymotic. The capillary system of the skin, and of the brain and its
membranes, is engorged, with, in some instances, serous
extravasations.
Prevention. This would demand the avoidance or correction of all
those conditions which contribute to tympany. In fevers and
extensive inflammations, when rumination is suspended, the diet
should be restricted in quantity and of materials that are easily
digested (well boiled gruels, bran mashes, pulped roots, etc.,) and all
bulky, fibrous and fermentescible articles must be proscribed. In
weak conditions in which tympany supervenes on every meal, a
careful diet may be supplemented by a course of tonics, carminatives
and antiseptics such as fœungrec oxide of iron, hyposulphite of soda
and common salt, equal parts, nux vomica 2 drs. to every 1 ℔. of the
mixture. Dose 1 oz. daily in the food, or ½ oz. may be given with each
meal.
Musty grain and fodder should be carefully avoided, also
mowburnt hay, an excess of green food to which the stock is
unaccustomed, clover after a moderate shower, or covered with dew
or hoarfrost, frosted beet, turnip, or potato tops, frosted potatoes,
turnips or apples, also ryegrass, millet, corn, vetches, peas with the
seeds fairly matured but not yet fully hardened. When these
conditions cannot be altogether avoided, the objectionable ration
should be allowed only in small amount at one time and in the case
of pasturage the stock should have a fair allowance of grain or other
dry feed just before they are turned out. Another precaution is to
keep the stock constantly in motion so that they can only take in
slowly and in small quantity the wet or otherwise dangerous aliment.
When it becomes necessary to make an extreme transition from
one ration to another, and especially from dry to green food,
measures should be taken to make the change slowly, by giving the
new food in small quantities at intervals, while the major portion of
the diet remains as before, until the fæces indicate that the
superadded aliment has passed through the alimentary canal.
Another method is to mix the dry and green aliments with a daily
increasing allowance of the latter. Some have avoided the morning
dew and danger of fermentation by cutting the ration for each
succeeding day the previous afternoon and keeping it in the interval
under cover.
Treatment. Various simple mechanical resorts are often effective
in dispelling the tympany. Walking the animal around will
sometimes lead to relaxation of the tension of the walls of the
demicanal and even to some restoration of the movements of the
rumen with more or less free eructation of gas. The dashing of a
bucket of cold water on the left side of the abdomen sometimes
produces a similar result. Active rubbing or even kneading of the left
flank will sometimes lead to free belching of gas. The same may be at
times secured by winding a rope several times spirally round the
belly and then twisting it tighter by the aid of a stick in one of its
median turns.
A very simple and efficient resort is to place in the mouth a block
of wood 2½ to 3 inches in diameter and secured by a rope carried
from each end and tied behind the horns or ears. This expedient
which is so effective in preventing or relieving dangerous tympany in
choking appears to act by inducing movements of mastication, and
sympathetic motions of the œsophagus, demicanal and rumen. It not
only determines free discharge of gas by the mouth, but it absolutely
prevents any accession of saliva or air to the stomach by rendering
deglutition difficult or impossible. A similar effect can be obtained
from forcible dragging on the tongue but it is difficult to keep this up
so as to have the requisite lasting effect. Still another resort is to
rouse eructation by the motions of a rope introduced into the fauces.
The passing of a hollow probang into the rumen is very effective as
it not only secures a channel for the immediate escape of the gas, but
it also stimulates the demicanal and rumen to a continuous
eructation and consequent relief. Friedberger and Fröhner advise
driving the animals into a bath of cold water.
Of medicinal agents applicable to gastric tympany the best are
stimulants, antiseptics and chemical antidotes. Among stimulants
the alkaline preparations of ammonia hold a very high place. These,
however, act not as stimulants alone, but also as antacids and
indirectly as antidotes since the alkaline reaction checks the acid
fermentation which determines the evolution of the gas. They also
unite with and condense the carbon dioxide. Three ounces of the
aromatic spirits of ammonia, one ounce of the crystalline
sesquicarbonate, or half an ounce of the strong aqua ammonia may
be given to an ox, in not less than a quart of cold water. Next to this
is the oil of turpentine 2 oz., to be given in oil, milk, or yolk of egg.
But this too is an antiferment. The same remark applies to oil of
peppermint (½ oz.), the carminative seeds and their oils, and the
stronger alcoholic drinks (1 quart). Sulphuric or nitrous ether (2 oz.)
may be given in place. Pepper and ginger are more purely stimulant
and less antiseptic. Other alkalies—carbonate of potash or soda, or
lime water may be given freely.
Among agents that act more exclusively as antiseptics may be
named: muriatic acid 1 to 1½ drs. largely diluted in water; carbolic
acid, creosote or creolin, 4 drs. largely diluted; sulphite, hyposulphite
or bisulphite of soda 1 oz.; kerosene oil ½ pint; chloride of lime 4
drs.; chlorine water 1 pint; wood tar 2 oz. The latter agent is a
common domestic remedy in some places being given wrapped in a
cabbage leaf, and causing the flank to flatten down in a very few
minutes as if by magic. The extraordinarily rapid action of various
antiseptics is the most conclusive answer to the claim that the
disorder is a pure paresis of the walls of the rumen. The affection is
far more commonly and fundamentally an active fermentation, and
is best checked by a powerful antiferment. Even chloride of sodium
(½ lb.) and above all hypochlorite of soda or lime (½ oz.) may be
given with advantage in many cases.
Among agents which condense the gasses may be named
ammonia, calcined magnesia, and milk of lime for carbon dioxide,
and chlorine water for hydrogen.
Among agents used to rouse the torpid rumen and alimentary
canal are eserine (ox 3 grs., sheep ½ gr. subcutem), pilocarpin (ox 2
grs., sheep ⅕ gr.), barium chloride (ox 15 grs., sheep 3 to 4 grs.),
tincture of colchicum (ox 3 to 4 drs.). Trasbot mentions lard or butter
(ox 4 oz., sheep ½ oz.), as in common use in France.
In the most urgent cases, however, relief must be obtained by
puncture of the rumen, as a moment’s delay may mean death. The
seat for such puncture is on the left side, at a point equidistant from
the outer angle of the ilium, the last rib and the transverse processes
of the lumbar vertebræ. Any part of the left flank might be adopted to
enter the rumen, but, if too low down, the instrument might plunge
into solid ingesta, which would hinder the exit of gas, and would
endanger the escape of irritant liquids into the peritoneal cavity. In
an extra high puncture there is less danger, though a traumatism of
the spleen is possible under certain conditions. The best instrument
for the purpose is a trochar and cannula of six inches long and ⅓ to
½ inch in diameter. (For sheep ¼ inch is ample.) This instrument,
held like a dagger, may be plunged at one blow through the walls of
the abdomen and rumen until stopped by the shield on the cannula.
The trochar is now withdrawn and the gas escapes with a prolonged
hiss. If the urgency of the case will permit, the skin may be first
incised with a lancet or pen knife, and the point of the instrument
having been placed on the abdominal muscles, it is driven home by a
blow of the opposite palm. In the absence of the trochar the puncture
may be successfully made with a pocket knife or a pair of scissors,
which should be kept in the wound to maintain the orifice in the
rumen in apposition with that in the abdominal wall, until a metal
tube or quill can be introduced and held in the orifices.
When the gas has escaped by this channel its further formation can
be checked by pouring one of the antiferments through the cannula
into the rumen.
When the formation of an excess of gas has ceased, and the
resumption of easy eructation bespeaks the absence of further
danger, the cannula may be withdrawn and the wound covered with
tar or collodion.
When the persistent formation of gas indicates the need of
expulsion of offensive fermentescible matters, a full dose of salts may
be administered. If the presence of firmly impacted masses can be
detected, they may sometimes be broken up by a stout steel rod
passed through the cannula. If the solid masses prove to be hair or
woolen balls, rumenotomy is the only feasible means of getting rid of
them.
In chronic tympany caused by structural diseases of the
œsophagus, mediastinal glands, stomach or intestines, permanent
relief can only be obtained by measures which will remove these
respective causes.
CHRONIC TYMPANY OF THE RUMEN.