Speech Analysis / Synthesis    

This demo is extracted from Professor Edward Lee's Home Page .


The rough structure of a speech synthesis model is shown below:
Voiced spectrum
The "noise" block generates what is called in the demonstrations below "hiss-type energy," while the "periodic pulses" block generates what is called "buzz-type energy." The period of the periodic pulses controls the "pitch". The audio is from an original Bell Labs recording of 1939. The speaker is alleged to be C. Voderson.


Introduction
The introduction to the vocoder itself has been processed by the vocoder, demonstrating reasonably good audio quality (by telephone standards, which emphasize intelligibility and speaker recognition over audio fidelity).

Comparison
Here, the vocoder output is compared to uncoded output (over a "public address system").

Unvoiced speech
Whispered speech is generated by setting the vocoder as if all speech were unvoiced (input to the synthesis filter is only "hiss-type energy"). Below, a plot of voiced speech (top) is compared to a plot of unvoiced speech (bottom) in the time domain.
Voiced/Unvoiced comparison
Notice that the voiced speech is much more periodic, while the unvoiced speech is much more random. The horizontal axis is "number of samples," and the sample rate is 8kHz. In the frequency domain, the spectrum of a segment of unvoiced speech will be smooth, as shown below:
Unvoiced spectrum
(This is not an actual plot of a spectrum, but rather a suggestive sketch.) By contrast, voiced speech, since it is roughly periodic, will have a more discrete spectrum, as shown below:
Voiced spectrum
For most languages, speech is fully intelligible in unvoiced (whispered) form. The voiced spectrum above has the same envelope (shown in red) as the unvoiced spectrum, so we would conclude that the linguistic information is the same in both signals. Indeed, only a few languages, like Mandarin Chinese, contain linguistic information in the periodicity of speech (Mandarin Chinese is "tonal," meaning that there is linguistic information in the inflections). The examples below will reinforce this point by altering the tonal information in interesting ways.

Voiced speech
Mechanical-sounding speech generated by setting the vocoder as if all speech were voiced (input to the synthesis filter is only "buzz-type energy").

Monotone speech
Here, both voiced and unvoiced sounds are produced, but the voiced sounds are held at a constant pitch, yielding a monotone effect.

Pitch modifications
Here, pitch is modified under the control of a hand dial.

One octave lower
An octave is a factor of two in frequency. In this demonstration, the vocoder halves the pitch of the speaker.

One octave higher
In this demonstration, the vocoder doubles the pitch of the speaker.

Inflection
"Inflection" is the variations in pitch in speech. The vocoder can be set to reduce or increase the inflection without shifting the pitch up or down.

Inflection manipulations on a song
In this demonstration, the inflection reduction and enhancement is demonstrated on a song.

Reversing the inflection
In this demonstration, inflection is reversed. That is, when the pitch of the original speech would be rising, here it is falling, and vice versa.

Special effects sounds
Here, the vocoder is used to synthesize non-speech sounds.

Vibrato
Vibrato is a musical term for a rapid fluctuation in pitch. This illustration uses the vocoder to introduce vibrato into a singing voice signal.

Jones family
In this demo, various of the above effects are combined to alter a single voice to play several roles in short skit.

Combining two voices
Here, a voice is shifted in pitch by a frequency interval known to musicians as a major third. The shifted voice signal is combined with the original to achieve a harmonious effect.

Combining three voices
Here, a voice is shifted in pitch by two frequency intervals to make what is known to musicians as a triad. The shifted voice signals are combined with the original to achieve a harmonious effect.

Permuting the frequency channels
Here, the three lowest frequency channels are redirected in synthesis to the three middle channels at higher frequencies. The result is a nasal effect, with the low frequencies missing.

Permuting the frequency channels
Here, the three middle frequency channels are redirected in synthesis to the three lowest channels. The result is a strange effect, with the middle frequencies missing.

Complete Audio File
The entire audio for the above demonstrations is available in Sun Audio format (.au files) (8,3700k).
Professor Edward Lee's Home Page.