Glossary
A mel spectrogram is the input format for modern speech recognition models. It's a 2D image where the X-axis is time, the Y-axis is frequency, and pixel brightness is amplitude — but with the frequency axis warped to the Mel scale, which is perceptually uniform (1 kHz feels half as high as 2 kHz to a human listener). Whisper, wav2vec, Conformer, and every current ASR model takes a mel spectrogram as input.
Why the Mel scale and not regular Hz?
Human hearing is logarithmic in frequency. The pitch difference between 100 Hz and 200 Hz sounds bigger than the difference between 5,000 Hz and 5,100 Hz, even though they're both 100 Hz apart numerically. The Mel scale is a non-linear transform that approximates how humans perceive pitch — equally-spaced points on the Mel scale sound equally far apart to us. Training ASR on mel spectrograms gives the model a perceptually-aligned representation, which generalizes better than raw frequency.
How a mel spectrogram is computed
Three steps. (1) Short-Time Fourier Transform (STFT): slide a 25 ms window across the audio in 10 ms hops, take the FFT of each window, get a regular linear-frequency spectrogram. (2) Mel filter bank: apply ~80 triangular filters spaced on the Mel scale, summing each filter's energy. (3) Log compression: take the log of each value to compress the dynamic range. Result: an 80-channel × T-frame matrix that's the mel spectrogram.
What it looks like
Open any spectrogram visualizer (Audacity, sonic visualiser) and load a vowel. You'll see horizontal bands at the formants — the resonances of the vocal tract. Different vowels ("ee" vs "oo") have different formant patterns. Consonants show up as bursts (plosives) or noise (fricatives). ASR models learn to read these patterns just like a phonetician would.
In practice
Inside Transcript.you's pipeline: when you upload an MP3, FFmpeg decodes it to PCM, we resample to 16 kHz, then compute the mel spectrogram (80 mel channels, 10 ms hop). That spectrogram is what gets sent to Whisper. The original audio waveform never reaches the model — only its mel spectrogram representation.
Related terms
Further reading
See AI transcription in action
Get started — freeLast updated: September 20, 2026
Frequently Asked Questions
What does Mel spectrogram mean in transcription?
A visual representation of audio frequencies plotted on the Mel scale (perceptually uniform). The standard input format for modern ASR neural networks, including Whisper.
Why does Mel spectrogram matter when choosing a transcription workflow?
Mel spectrogram affects how you understand transcript quality, timing, compatibility, or the technology behind speech-to-text results.
Where can I apply Mel spectrogram on Transcript.you?
Use the related workflow at /ai-transcription when you want to see how this glossary concept connects to an actual transcription task.
Browse the full transcription glossary or read the complete guide to AI transcription.