# Generate a Mel Spectrogram PNG
Render a local audio file as a PNG mel spectrogram with ML-style FFT, hop, mel-band, dB, colormap, and image-size controls.
## Run it
- **CLI:** `gizza tool mel-spectrogram-generator 'url=https://example.com/input' 'input_format=base64' 'n_fft=2048' 'hop_length=0' 'n_mels=128' 'fmin=0' 'fmax=0' 'mel_scale=slaney' 'window=hann' 'scale=db' 'top_db=80' 'db_reference=peak' 'colormap=magma' 'width=1000' 'height=400' 'channel=mix' 'center=true' 'resample_hz=0'`
- **Web:** https://gizza.ai/tools/mel-spectrogram-generator/
- **Agents:** machine-readable descriptor (parameters JSON Schema) at https://gizza.ai/tools/mel-spectrogram-generator/tool.json
## Inputs
- `input` — Audio file _(file; accept: audio/*)_
- `input_format` — Input encoding _(field)_
- `n_fft` — FFT / window size (samples) _(field)_
- `hop_length` — Hop length (samples, 0 = n_fft/4) _(field)_
- `n_mels` — Mel bands _(field)_
- `fmin` — Low frequency (Hz) _(field)_
- `fmax` — High frequency (Hz, 0 = Nyquist) _(field)_
- `mel_scale` — Mel scale _(field)_
- `window` — Analysis window _(field)_
- `scale` — Intensity scale _(field)_
- `top_db` — dB dynamic range _(field)_
- `db_reference` — dB reference _(field)_
- `colormap` — Colormap _(field)_
- `width` — Output width (px, 0 = natural) _(field)_
- `height` — Output height (px, 0 = natural) _(field)_
- `channel` — Channel _(field)_
- `center` — Center frames (librosa-style padding) _(field)_
- `resample_hz` — Analysis sample rate (Hz, 0 = native) _(field)_
## Output
- Mel spectrogram (image)
## Query parameters
Open the tool pre-filled and auto-run via URL:
- `input_format` — Input encoding
- `n_fft` — FFT / window size (samples)
- `hop_length` — Hop length (samples, 0 = n_fft/4)
- `n_mels` — Mel bands
- `fmin` — Low frequency (Hz)
- `fmax` — High frequency (Hz, 0 = Nyquist)
- `mel_scale` — Mel scale
- `window` — Analysis window
- `scale` — Intensity scale
- `top_db` — dB dynamic range
- `db_reference` — dB reference
- `colormap` — Colormap
- `width` — Output width (px, 0 = natural)
- `height` — Output height (px, 0 = natural)
- `channel` — Channel
- `center` — Center frames (librosa-style padding)
- `resample_hz` — Analysis sample rate (Hz, 0 = native)
- `url` — fetch the input file from a public URL (CORS-permitting)
Example: `https://gizza.ai/tools/mel-spectrogram-generator/?url=https://example.com/input&input_format=base64&n_fft=2048&hop_length=0&n_mels=128&fmin=0&fmax=0&mel_scale=slaney&window=hann&scale=db&top_db=80&db_reference=peak&colormap=magma&width=1000&height=400&channel=mix¢er=true&resample_hz=0`
---
## About this tool
Mel spectrograms compress an audio clip into the time-frequency view used by many speech, music, keyword-spotting, and audio-classification pipelines. This tool decodes the audio locally, applies an STFT, maps FFT bins through triangular mel filters, scales the energy, and renders the result as a PNG image.
Worked example: upload a WAV voice clip, choose the “Speech 16 kHz compact” preset, and download the PNG. The output uses 80 mel bands, a 1024-sample FFT, a 256-sample hop, an 8 kHz high-frequency edge, and a 16 kHz analysis sample rate — a compact diagnostic view for speech models.
Limits: input audio is capped at 24 MiB and 4,000,000 decoded samples. Very short clips may need `center=true` or a smaller FFT. `width=0` and `height=0` use the natural matrix size: one pixel per frame and one row per mel band.
## FAQ
Is this the same as an MFCC extractor?
No. A mel spectrogram keeps the mel-band energy image before the DCT step. MFCC tools convert the log-mel bands into cepstral coefficients for tabular features; this tool renders the mel bands directly as a PNG for inspection or image-based pipelines.
Which settings match common audio-ML examples?
A common starting point is `n_fft=2048`, `hop_length=512`, `n_mels=128`, `mel_scale=slaney`, `scale=db`, and `center=true`, which mirrors the defaults many librosa-style examples use. Speech pipelines often resample to 16 kHz and use 64 or 80 mel bands.
Why are low frequencies at the bottom of the PNG?
That orientation matches the way spectrograms are usually read: time runs left to right and frequency rises from bottom to top. The summary reports the loudest mel-band frequency so you can sanity-check tones and hums.
What do `peak` and `full_scale` dB references change?
`peak` maps the loudest cell in the clip to the brightest color, which is useful for visual contrast. `full_scale` keeps the reference tied to digital full scale, so quiet files stay visibly dimmer and are easier to compare across clips.
## Related tools
- [Export Audio Features to CSV](https://gizza.ai/tools/audio-feature-csv-exporter/): Export per-frame audio features — RMS, spectral centroid, ZCR, rolloff, flatness, bandwidth, flux — to CSV, TSV, or JSON. Runs in your browser.
- [Extract MFCC Features from Audio](https://gizza.ai/tools/mfcc-extractor/): Extract MFCC features from audio as CSV, TSV, or JSON. Tune speech defaults, librosa-style settings, deltas, and sample-rate options locally.
- [Waveform Image from Audio](https://gizza.ai/tools/waveform-image/): Turn any audio file into a waveform PNG in your browser — pick the size, solid or gradient wave colors and background, or keep it transparent. Nothing is uploaded.
- [Convert AIFF to FLAC](https://gizza.ai/tools/aiff-to-flac/): Convert AIFF or AIF audio to lossless FLAC in your browser — preserve samples and textual tags, choose FLAC compression level 0-12, and download the result.
- [Convert AIFF to WAV](https://gizza.ai/tools/aiff-to-wav/): Free AIFF to WAV converter that runs in your browser. Choose the PCM bit depth, sample rate and channel layout — at a matching depth the samples are identical.