10

This paper may be useful for its collection of references. This paper is a single simple read. The problem is that the spectral properties of fricatives are usually reported in terms of the frequency of the spectral peak, where /f/ and /s/ are clearly different but also above the cutoff frequency for the phone. Another part of the black box that you're up ...


8

why the harmonic frequencies are integer multiples of the fundamental frequency? Physics! Pretty much all of the physical processes that create vibrations at a base frequency, also create vibrations at integer multiples of that base frequency. This includes the way the human vocal chords work. You can artificially synthesize a sound with any frequencies ...


6

As @jlawler mentioned in his comment, word-final [f] and [s] are likely to be confused over the phone in general. The place of articulation of a fricative is mainly cued by two things--(1) the center of gravity (or diffuseness) of its noise and (2) the direction of the formant transitions going into and coming out of it. In the case of word-final ...


6

Broadly, you're describing the entire (not-fully-solved) problem of automatic speech recognition/automatic transcription. However, if you have the text of the sentences (e.g., if the recordings are scripted, or if you've manually transcribed their speech), then the problem is more tractable: you want 'forced alignment'. A popular software option for that is ...


6

Contrary to the expectations of some commentators, doctor-patient corpora are available (under some conditions, needing to sign some licence and confidentially agreement) for research. The standard entry point for a search for such corpora is the CLARIN Virtual Language Observatory and entering doctor patient in the search slit gives currently twelve results....


5

https://web.archive.org/web/20171130080859/https://hacks.mozilla.org/2017/11/a-journey-to-10-word-error-rate/ : Our word error rate on LibriSpeech’s test-clean set is 6.5%, which not only achieves our initial goal, but gets us close to human level performance. [...] (5.83% according to the Deep Speech 2 paper). On a MacBook Pro, using the GPU, the model ...


5

There is no such thing. What can be recorded (i.e. what occurs in speech) is an allophone, and phonemes are abstractions built from allophones. Every allophone occurs in a particular context, so the acoustic value of /k/ in "keep" is different from the value of /k/ in "cool". Allophones also do not have well-defined acoustic beginning and ending points; for ...


4

Phones are a "thing" because they were the first decent method of objectively and accurately recording unwritten languages (in the 19th century). Back then, if you heard a Lushootseed speaker translate English "The bear ate the salmon", the standard practice was to guess with untrained English-speaker ears that the person said "Oo uhshluh tube tea skuchicuss ...


4

Assuming the goal of writing a speech recognition program that does what the human mind does, a large non-linguistic front end must be dealt with first (a front end that is decidedly not part of linguistics), namely, how does the human auditory system work, starting at the ear? What exactly hits the brain coming from the primary auditory cortex, or later? If ...


4

The other answers have hit the highlights, going so far as to suggest that it is impossible in principle. Contrariwise, I argue that it could be done in principle, as long as you don't overstate what IPA does. IPA is a conventional system (conventions can be voted in or out) for grouping a range of acoustic events to letters, with a coarse enough granularity ...


4

Praat is the main program used to analyze sound data for phonetics research. It's available for free at the link. You can use the program to add markers and replay snippets, as well as analyze formants.


4

This suggests a possible meta-study on intelligibility of technical works by native and non-native speakers. A technical paper in phonology might be unintelligible because of the linguistic structure of the article, or because of the subject matter. The same is true in speech and hearing science, and various other areas. Fortunately, SPHS and phonetics ...


3

I'm pretty sure the short answer for this is: "there's no algorithm that will always agree with your judgements about what is/isn't a vowel/consonant (cf. glides) and will always agree with your judgements about where said phones start/end in the signal" BUT, a bit of quick Googling turned up the following, which may be useful if you're familiar with ...


3

Kaldi provides WER of 4.28% whereas deepspeech gives 5.83% on librispeech clean data. Check this out: https://github.com/syhw/wer_are_we


3

Take a look at 2Hz Noise Suppression API. It's language-agnostic and is a REST API so you can use it from any programming language. Currently the API isn't designed for real time however this will be added down the road. (full disclosure - I work at 2Hz)


3

IMO this is a classical terminological question. In some views, by definition it is an "allophone". The phones are the detailed surface sounds, a symbolic representation of speech without prejudice to what language the sound occurs in. Thus, a very narrow phonetic transcription. Given that [k, kʰ, kʲ, kʲʰ] etc. may all occur in complementary environments and ...


3

This is impossible in principle, at least in the form that you asked. A modified version of the task might be possible. The main reason is that the input would be an acoustic waveform, which needs to be parsed into discrete segmental chunks, then labeled according to a standard. The letters and articulatory descriptions are standardized and easily accessible,...


2

Without question, it is an order of magnitude more difficult to implement speech recognition. You can do passable speech synthesis knowing nothing about context, grammar, etc. You will get some issues with homographs but other than that you will get recognizable speech. Now, it will be greatly lacking in prosody and stress, so you need to do some language ...


2

If you know beforehand the content of the speech signal but you just don't know which parts of the waveform correspond to which parts of the utterance, then your problem is reduced to an alignment problem--much easier to deal with than a full-fledged speech recognition (plus alignment) problem. You already know what language it is, and you know what ...


2

I suspect you are not using the expression "phone level" in its technical sense. Also "detecting words at the phone level" isn't meaningful, but perhaps you intend "detecting words, given a waveform-to-phone conversion". To clarify, a "phone" is a reduction of the continuous acoustic waveform to percepually-based units (i.e. a narrow phonetic transcription). ...


2

Human speech is noisy, and speech recognition must be able to find patterns in the noise. Phones have a series of articulatory attributes: places and manners of articulation, tongue shape, etc.; which cause voice resonance and distortion. All of those variables are continuous, and a continuous change in one of these parameters produces a continuous change ...


2

Researchers speak of Segmental Variability or Density of Phonetic Encoding in this respect, see, e.g., Zofia Malisz, Erika Brandt, Bernd Möbius, Yoon Mi Oh, and Bistra Andreeva,Dimensions of Segmental Variability: Interaction of Prosody and Surprisal in Six Languages


2

Removing noise usually corrupts the speech and harms the performance. It is frequently more accurate to decode noisy speech than noise-cleaned speech, in particular because recognizer does noise cleaning or noise compensation by itself. For other cases except speech recognition, rnnnoise is reasonable.


2

Modern data-driven approaches to speech recognition are ignorant about but robust to this one of many aspects of orthography that are not fully phonetic but occur relatively consistently in the training data. For one, it is more an issue of mapping n:1 (different realisations of one set of characters), whereas challenges for speech recognition generally ...


2

In my experience, speech-recognition systems do quite a bit of preprocessing on the signal before trying to interpret it; part of that is getting rid of the noise. The key is, most speech signals are periodic: the vocal folds generate a sawtooth-esque wave, and the vocal tract then applies a filter to it. So all voiced sounds(*) have a very predictable ...


2

You will probably have to dig elsewhere to find actual data on lateral lisps. One large caveat is that you can't compare children and adults, and the data on characteristic properties of phonemes will come from adults whereas lateral list is a condition to be eradicated with young children. You can however read this paper, which compares acoustic properties ...


2

There is no program that converts samples speech to IPA. Some program might appear to do that, by converting an utterance of language Z into its orthographic form then following conventional text-to-IPA rules to yield a psuedo-trancription, which that would not detect variations (among native speakers) between [i] and [ɛ] as the first vowel of "economic&...


1

I believe phonologists often don't make a distinction between allophonic and non-allophonic variation in the realization of phonemes, but, when they do, it goes something like this: allophonic variation is discrete, driven by a distinction that can be captured as a simple rule, maps to a plausible (and usually independently attested) binary distinction in ...


1

Can we divide phonemes which are specific to a language into some more fundamental entities same across all languages ? There is the spectrogram (https://en.wikipedia.org/wiki/Spectrogram), maybe that is more fundamental than you wanted. The raw datasets are just .wav or .mp3 files. None of that is specific to language. The language-specific part is ...


1

If you actually try to speak a very long Turkish word, for example Afyonkarahisarlılaştırabildiklerimizdenmişsinizcesine, is it recognised? Today's speech recognition is end-to-end, that is, the input in training is just the audio and the transcript. There is ideally nothing language-specific in that part. It can be done, for example, with seq2seq ...


Only top voted, non community-wiki answers of a minimum length are eligible