Here are several examples of speech and speech coding. The source speech waveform is bandlimited and sampled at 8 kHz. Each sample is represented using 12 bits (uniform quantization). The speech is encoded using two different methods. The first method is variable- rate coded-excited linear prediction (VCELP). This approach partitions the sampled speech into frames (of typically 20 ms) and each frame is encoded at a rate of 8, 4, 2, or 1 kbps. The encoding rate to be used for each speech frame is determined by an "activity measure" selected to monitor how much energy occurs in the frame. This is useful in two-way communications, since rarely are both talkers speaking simultaneously, so that while one talker is speaking, the rate used to encode the sound from the other talker can be kept to a minimum. This algorithm is used in some cellular telephone systems. The second method is called improved multiband excitation (IMBE) speech coding. In this approach the speech is also partitioned into frames, and each frame is analyzed using the DFT to determine pitch and harmonic frequencies. The magnitude of the amplitude spectrum is then (coarsely) quantized and encoded and the phase is not encoded. The decoding involves synthesizing sinusoids for the encoded frequencies and amplitudes, and carefully maintaining continuity of phase between one frame and the next.

Original Audio (48 kHz sampling rate)

Retain 20% of DFT coefficients

Retain 10%

Retain 5%

Original speech (sent1.wav)

Speech sent1 subsampled by two (with aliasing)

Speech sent1 subsampled by four (with aliasing)

Speech sent1 subsampled by eight (with aliasing)

Original speech (sent3.wav)

Original speech (sent4.wav)

Sent4 truncated to 8016 samples (sent4-8016samples.dat)

IMBE Encoded at 4.8 kbps (sent1)

IMBE Encoded at 4.8 kbps (sent4)

Variable-rate CELP Encoded at 8 kbps (sent1)

Variable-rate CELP Encoded at 8 kbps (sent4)

Speech with echo (sampled at 8 kbps)

Speech with 100 ms echo (sampled at 8 kbps)