Representing Acoustics with Mel Frequency Cepstral Coefficients Lecture 7 Spoken Language Processing Prof. Andrew Rosenberg
Representing Acoustic Information 16-bit samples 44.1kHz sampling rate ~86kB/sec ~5MB/min Waves repeat -- Much of this data is redundant. A good representation of speech (for recognition) Keeps all of the information to discriminate between phones Is Compact. i.e. Gets rid of everything else
Frame Based analysis Using a short window of analysis, analyze the wave form every 10ms (or other analysis rate) Usually performed with overlapping windows. e.g. FFT and Spectrogram
Overlapping frames Spectrograms allow for visual inspection of spectral information. We are looking for a compact, numerical representation 10ms 10ms 10ms 10ms 10ms
Example Spectrogram Example Spectrogram from Praat
Standard Representation in the field Mel Frequency Cepstral Coefficients MFCC Pre-Emphasis window FFT Mel-Filter Bank energy log 12 MFCC 12 ∆ MFCC 12∆∆ MFCC 1 energy 1 ∆ energy 1 ∆∆ energy Deltas 12 MFCC FFT-1
Pre-emphasis Looking at spectrum for voiced segments, there is more energy at the lower frequencies than higher frequencies. Boosting high frequencies helps make the high frequency information more available. First-order high-pass filter for pre-emphasis. Figure 9.9
Windowing Overlapping windows allow analysis centered at a frame point, while using more information. Figure 9.10
Hamming Windowing Discontinuities at the edge of the window can cause problems for the FFT Hamming window smoothes-out the edges. Figure 9.11, Figure 9.12
Hamming Windowing Discontinuities at the edge of the window can cause problems for the FFT Hamming window smoothes-out the edges. Figure 9.11, Figure 9.12
Discrete Fourier Transform The algorithm for calculating the Discrete Fourier Transform (DFT) is the Fast Fourier Transform. http://clas.mq.edu.au/acoustics/speech_spectra/fft_lpc_settings.html Australian male /i:/ from “heed” FFT analysis window 12.8ms
Mel Filter Bank and Log Human hearing is not equally sensitive at all frequency regions. Modeling human hearing sensitivity helps phone recognition. MFCC approach: Warp frequencies from Hz to Mel frequency scale. Mel: pairs of sounds that are perceptually equidistant in pitch are separated by an equal number of mels.
Mel frequency Filter bank Create a bank of filters collecting energy from each frequency band, 10 filters linearly spaced below 1000Hz, logarithmic spread over 1000Hz. Figure 9.13
Cepstrum Separation of source and filter. Source differences are speaker dependent Filter differences are phone dependent. Cepstrum is the “Spectrum of the Log of the Spectrum” – inverse DFT of the log magnitude of the DFT of the signal
Cepstrum Visualization Peak at 120 samples represents the glottal pulse, corresponding to the F0 Large values closer to zero correspond to vocal tract filter (tongue position, jaw opening, etc.) Common to take the first12 coefficients Figure 9.14
Deltas and Energy Energy within a frame is just the sum of the power of the samples. The spectrum of some phones change over time – the stop closure to stop burst, or slope of a formant. Taking the delta or velocity and double delta or acceleration incorporates this information
Summary: MFCC Commonly MFCCs have 39 Features 39 MFCC Features 12 Cepstral Coefficients Delta Cepstral Coefficients Delta Delta Cepstral Coefficieints 1 Energy Coefficients Delta Energy Coefficients Delta Delta Energy Coefficients
Next Class Introduction to Statistical Modeling and Classification Reading: J&M 9.4, optional 6.6