Speak and sing - introduction

Speech scaling and pitch correction

The speech scaling and pitch correction program, or “Speak&Sing”, generates a properly-timed and pitch-accurate sample of a known song from recorded spoken words.

A voice modulation application which detects the timing and pitch of the recorded input and automatically performs time scaling and pitch correction to match the speech to a pre-selected song, producing a musical output.

Making a smarter autotuner

Pitch correction is used by musicians to improve the sound of song vocals by fixing off-key singing or adding distortion. It can be applied real-time using a synthesizer keyboard or added after recording. However, these “autotuners” can’t fix off-tempo singing, and automatic autotuners depend on the singer to be relatively close to the right pitch.

Goals for the Speak and Sing:

• Proof-of-concept of automated syllable detection, time scaling, and pitch correction in one robust application
• Provides open-ended, customizable options for audio processing
• Demonstrates time and frequency DSP applications using MatLab

Implementation: an overview

Recording:

Input voice samples are recorded in mono-channel audio at a sampling frequency of 16,000 (although any sampling frequency can be used). It is then imported into MatLab and the following functions are run in sequence:

Song Interpretation and Retrieval:

Contains data for selected songs based on sheet music. It returns a vector of fundamental note frequencies and note lengths depending on the song selected and the desired tempo.

Available Songs:

1. Christina Aguilera - Genie in a Bottle

2. Mary Had a Little Lamb

3. Row, Row, Row Your Boat

Syllable Detection

Analyzes the input speech data and determines the locations and lengths of each syllable. After dividing the signal up into several short windowed pieces, it detects the periodicity and energy of each window to determine the type of sound (vowel, consonant, or noise). The locations of the syllables is then determined based on the pattern of sounds.

Time scaling

Interprets the detected syllable locations and stretches or shrinks the syllable to match the length of the word in the song. The time scaling is performed using a time-domain Waveform Similarity Overlap Add (WSOLA) algorithm, which breaks up the signal into overlapping windows and copies each window to a new location, either closer together or further apart. This stretches or compresses the length of the speech without losing quality or information.

Pitch correction

Detects the pitch of the signal, compares it to the desired pitch of the song, and makes pitch corrections. Pitch detection is done using FAST-autocorrelation, in which small windows of the signal are offset and autocorrelated to find the period, and thus frequency, of the signal for that interval. Pitch correction is performed with Pitch Synchronous Overlap Add (PSOLA), which moves windowed segments closer or further apart and overlap-added to alter the frequency without loss of sound.

The result

The resulting audio file sounds like the input speech or singing, but the words will now line up with those of the original song and the pitch will be adjusted. The resulting impact on sound is that the recorded input will now sound more like the song, without compromising the original voice or speech.