Mel-frequency cepstral coefficients (MFCCs) are a compact set of features that summarize the shape of a sound's spectrum. For decades they were the dominant input to speech recognition systems. Derived from a mel-spectrogram, they capture the vocal-tract characteristics that separate one speech sound from another using only about a dozen numbers per audio frame.
A brief history of MFCCs
MFCCs come from a 1980 paper by Steven Davis and Paul Mermelstein, who compared several representations for recognizing spoken syllables and found that mel-frequency cepstral coefficients worked best. The design borrows from the mel scale, which spaces frequency the way human hearing does, packing more resolution into the lower frequencies where speech carries most of its information.
Through the 1990s and 2000s, MFCCs became the near-universal front end for speech recognition.
How MFCCs work
Computing MFCCs adds one step on top of a mel-spectrogram. You take the log of the mel-filterbank energies, then apply a discrete cosine transform and keep only the first coefficients, typically around 13.
The cosine transform decorrelates the values and concentrates the useful information into those low coefficients, which describe the overall spectral envelope, the part that signals which phoneme was spoken, while discarding fine detail. This is the idea behind the cepstrum, loosely the spectrum of a spectrum.
MFCCs in classical versus modern systems
MFCCs were the feature for hidden Markov model systems and early neural recognizers, chosen because they were small and worked well with the limited compute of the era. Modern end-to-end models increasingly take a mel-spectrogram, or even raw audio, and learn their own features internally, so MFCCs are less central than they once were.
MFCCs remain common in embedded and on-device systems, speaker recognition, and plenty of production pipelines that predate the deep-learning shift.
Why MFCCs matter
You will still meet MFCCs in older systems, compact on-device recognizers, and feature extraction documentation, so the term is worth knowing even as the field moves toward learned features. Understanding how they are built also demystifies the acoustic model's input stage and clarifies why "features" in speech usually means a small per-frame vector rather than the raw sound.
Frequently asked questions
What is MFCC? MFCC stands for mel-frequency cepstral coefficients, a compact set of numbers describing the spectral shape of a short slice of audio. They were the standard input feature for speech recognition for decades.
What does MFCC stand for? Mel-frequency cepstral coefficients. "Mel" refers to the perceptual frequency scale, and "cepstral" refers to the cepstrum, the transform applied to the log spectrum to produce them.
What is MFCC in audio processing? It is a feature-extraction method that converts a waveform into roughly 13 coefficients per frame, capturing vocal-tract shape for tasks like speech and speaker recognition.
Are MFCCs still used? Yes, in on-device recognizers, speaker verification, and older or resource-limited pipelines. Many modern end-to-end models skip them and learn features from a mel-spectrogram or raw audio instead.