Spectrogram review

spectrograms are basic tools for speech and audio analysis.

• formant: formant are the resonant frequencies created within speech.
• speech elements:
• vowels: voiced sounds, such as a, e, etc..
• fricatives: fricatives are the phonemes that are produced by a constriction in the vocal tract. They don't usually contain much resonant frequencies (formants), but have content across frequency spectrum.
• plosives: plosives are transient bursts created by closure of vocal tract followed by release, which can be voiced or unvoiced.
• diphthongs / glides: glides are characterised by spectral movement of formants over time.
• nasals: nasals are resonant sounds produced by vibration within the nasal cavity.
• defining these elements in a spectrogram:
• formant: horizontal bands
• fricatives: vertical bands of flat spectrum 'noise'
• vowels: several formant over long period
• glides / diphthongs: moving formant; gradual vertical movement of a formant horizontal band
• plosives: stop in spectrum before vowels etc.
• nasals: two or more formants usually with a fairly large gap in between where there is a missing formant; just as vowels but with 'hole in lower spectrum'
• signal processing stages to produce a spectrogram:
• signal segmentation and windowing
• transformation to frequency domain via DFT with zero padding
• Log magnitude spectrum and stacking vectors in a matrix
• magnitude to colour mapping and display
• four main parameters / choices in producing a spectrogram:
• sampling rate
• DFT length
• segmentation window length / zero padding length
• overlap
• spectral resolution
• narrowband (long window) spectrogram makes harmonic structure clear
• bandwidth of 45~50Hz, e.g. Fs=44.1KHz, FFT size should be ~1024
• associated with glottal source
• wideband (short window) spectrogram makes formant structure clear
• bandwidth of 300~500Hz, e.g. Fs=44.1KHz, FFT size should be ~128
• dark formant bands that change with vowels not pitch
• formants associated 'filter properties' of vocal tract above larynx

Free Template Blogger collection template Hot Deals BERITA_wongANteng SEO

Automatic Speech Recognition (ASR) review

1、Problem Formulation: Frequency representations need to be invariant to pitch changes
建立一个ASR系统，必须考虑的要素有：
- timing variation
- loud / quiet speech
- speaker effects, such as gender, accents and vocal mannerisms
- contextual effects

2、speech feature extraction
① speech feature means the compact representations that can highlight distinguishing information extracted from the audio signal.
② source-filter theorysource is the excitation signal, such as oscillation of the glottis;filter is the effect of the time varying vocal tract;a speech signal can be considered as the convolution of the source and the filter.
③ a typical speech encoder: how source and filter are estimated
source can be generated using a white noise generator for an unvoiced sound or a pitch detector / plus generator for a voiced sound.filter can be estimated using an LPC filter or a suitably defined filter-bank.
④ cepstrum analysis: to separate the source and filter elements of speech using spectral methods:
     idft (log |dft (s(t))|)
⑤ LPC featuresall pole parameter estimation

3、Linguistic categories for speech recognition
-phone and phoneme
-IPA
-allophone: in phonetics, an allophone is one of a set of multiple possible spoken sounds (or phones) used to pronounce a single phoneme. For example, [pʰ] (as in pin) and [p] (as in cap) are allophones for the phoneme /p/ in the English language. Speakers treat them as the same phones, but they can be pronounced differently.

4、statistical sequence recognition: Hidden Markov Models
Free Template Blogger collection template Hot Deals BERITA_wongANteng SEO