Chapter_1_4007

Multimedia Networking - Digital Audio

• Sound is a wave phenomenon akin to light, but macroscopic, involving air molecule compression and expansion due to a physical device (e.g., a speaker).
• Sound, as a pressure wave, has continuous values, unlike digital representations.
• Sound waves exhibit ordinary wave behaviors such as reflection, refraction, and diffraction.
• Digitization converts audio waves into a stream of numbers, ideally integers, for efficiency.

Digitization

• An analog signal is a continuous measurement of a pressure wave.
• Digitization requires sampling in both time and amplitude.
• Sampling measures a quantity at evenly spaced intervals.
• Sampling frequency determines the rate of sampling (e.g., 8 kHz to 48 kHz for audio).
• Quantization is sampling in the amplitude dimension.
• Quantization involves representing amplitudes by certain values (steps), and rounding introduces inexactness (quantization noise).

Nyquist Theorem

• The Nyquist theorem dictates the sampling frequency required to reproduce the original sound accurately, ideally at least twice the maximum frequency in the signal (Nyquist rate).
• A sampling rate equal to the actual frequency results in a false signal (a constant with zero frequency).
• Sampling at 1.5 times the actual frequency yields an incorrect frequency (aliased frequency) lower than the correct one, with a doubled wavelength.
• If a signal is band-limited (with a lower and upper frequency limit), the sampling rate must be at least twice the highest frequency component (Nyquist rate = 2 * fmax).
• Nyquist frequency is half the Nyquist rate.

Signal-to-Quantization Noise Ratio (SQNR)

• Quantization noise results from the conversion of continuous values into discrete values.
• SQNR quantifies the quality of quantization, defined as the signal power to the quantization noise power (in decibels).

Signal-to-Noise Ratio (SNR)

• SNR measures the signal quality by comparing the power of the correct signal and noise.
• Measured in decibels (dB): 1 dB is a tenth of a bel.
• Defined using base-10 logarithms of squared voltages (SNR = 20 log₁₀(Vsignal / Vnoise).
• This is proportional to the square of the voltage. For example, if signal voltage is 10x noise, then SNR = 20 dB.

Linear vs. Non-linear Quantization

• Linear format stores samples as uniformly quantized values.
• Non-uniform quantization sets up more finely-spaced levels where human hearing acuity is highest.
• Weber's Law states that equally perceived differences in response have values proportional to the absolute levels of the stimulus.
• Non-linear quantization transforms an analog signal from the raw space to a theoretical space and uniformly quantizes the results.

Audio Quality vs. Data Rate

• Uncompressed data rate increases with more bits used for quantization, and stereo doubles the bandwidth.
• Examples of audio qualities and their corresponding parameters are given in a table (e.g., telephone quality, CD quality).

Coding of Audio

• Involves quantization and transformation of data.
• Temporal redundancy in audio exploits the differences between consecutive signals (reducing size of values) to enable more likely values using shorter bit lengths.
• PCM (Pulse Code Modulation) is a general term for a process involving producing quantized sampled output, with variations such as DPCM (differences), DM (crude, efficient), and ADPCM (adaptive variant).

Pulse Code Modulation (PCM)

• Given a bandwidth for speech (50 Hz to 10 kHz), the Nyquist rate dictates a sampling rate of 20 kHz.
• With 8 bits per sample, the bit rate for mono speech is 160 kbps.
• Standard telephony uses a bit rate of 64 kbps, considering speech signal max frequency of 4 kHz. High/low frequencies are removed using band-limiting filters.

Differential Coding of Audio

• Stored as differences in simple PCM, using fewer bits. This can assign short codes to common differences and long codes to infrequent ones.

Lossless Predictive Coding

• Predictive coding transmits differences between successive samples instead of samples themselves.
• Predicting the next sample as equal to the current sample. The error (difference) is transmitted.
• Some function of previous values can be used to gain a better prediction.
• Linear predictor function is typically employed.

Differential PCM (DPCM)

• DPCM is similar to predictive coding but now includes a quantizer step.

Distortion

• Distortion is the average squared error between predicted and actual signal values.
• D = (1/N) Σ(fn - ˜fn)²

Adaptive DPCM (ADPCM)

• ADPCM adapts the coder to better suit input signals, changing step size and decision boundaries.
• Adapts the predictor and quantizer. Methods exist employing forward and backward adaptive quantization.

Vocoders

• Algorithms for speech synthesis using limited bit-rates.
• Techniques such as using a modal speech waveform in time (LPC) or breaking down into frequency components (channel vocoder/formant vocoder) exist, for modelling salient or important frequencies.
• Not as good of a simulation for natural speech.

Phase Insensitivity in Speech

• Perceptual quality of speech doesn't depend on precise phase reconstruction of the waveforms, rather the amount of energy produced. Examples are shown.

Channel Vocoder

• Operates at low bit-rates (1-2 kbps), by filtering the signal into frequency components and determining their power levels, analyzing pitch, and using excitation (voiced or unvoiced). Diagram is included for reference.

Formant Vocoder

• Recognizes and encodes the important peaks (formants) in speech signals to produce understandable audio at low bit-rates (1 k bps) .

Linear Predictive Coding (LPC)

• Extract salient features of speech directly from the waveform and uses a time-varying model to produce speech (using equations to find the vocal tract coefficients);
• LPC features include speech parameter transmissions, not actual signals using small bit rates.

LPC Coding Process

• Process involves determining whether the segment is voiced or unvoiced to select the generator type (wideband-noise or pulse train).

Code Excited Linear Prediction (CELP)

• Attempts to improve on LPC by using a codebook of excitation vectors. More complex than LPC, providing similar audio quality at higher bitrates.

Algebraic Code Excited Prediction (ACELP)

• Distributes pulses as excitation for the linear prediction filter, allowing for large codebooks and reducing processing/storage needs.
• Using for different G series standard applications.

Adaptive Multi-Rate (AMR)

• Speech coding optimized for link conditions and offers several bit rates.
• Using discontinued transmission, voice activity detection, and comfort noise generation to reduce bandwidth use during silence periods.
• Sampling frequency and bit rates described.

Voice Activity Detection (VAD)

• Algorithm for speech detection from audio samples, used in speech coding and recognition.
• Can determine whether speech is present or absent, and if present, whether voiced or unvoiced.

Discontinuous Transmission (DTX)

• Momentarily powers down or muting mobile devices during pauses in speech transmission.
• The usage of this technique conserves battery life, eases on components, and reduces interference.

Comfort Noise

• Artificial background noise added to fill silences resulted from the voice activity detection.
• Prevents the other end from assuming a cut transmission, based on volume levels.

Adaptive Multi-Rate - Wideband (AMR-WB)

• Built on adaptive multi-rate technology (AMR).
• Utilizes various bit rates.

Cisco VoIP Implementations

• VoIP network benefits - e.g., efficient bandwidth use, lower transmission costs, improved employee productivity.
• Describes different VoIP network components (MCU, Application Servers, call agents) and describes their interactions.
• Analog to IP network conversion necessary for legacy systems.

Lossless and Lossy Compression

• Lossless methods produce identical output; lossy methods do produce an approximation to the original.
• Quantization is a main source of loss. 

Test your knowledge on the concepts of information theory and predictive coding, including key features like lossless predictive coding, differential PCM, and ADPCM. Answer questions about the fundamental principles that govern compression algorithms and the relationship between probability and information.