{ "slug": "mfcc-extractor", "name": "gizza-ai/mfcc-extractor", "version": "0.1.0", "title": "MFCC Extractor — gizza.ai", "description": "Extract MFCC features from audio as CSV, TSV, or JSON. Tune speech defaults, librosa-style settings, deltas, and sample-rate options locally.", "tags": [ "audio", "mfcc", "speech", "features", "cepstrum", "csv", "machine-learning" ], "category": "audio", "urls": { "page": "https://gizza.ai/tools/mfcc-extractor/", "markdown": "https://gizza.ai/tools/mfcc-extractor/index.md", "descriptor": "https://gizza.ai/tools/mfcc-extractor/tool.json", "deep_link_example": "https://gizza.ai/tools/mfcc-extractor/?url=https://example.com/input&input_format=base64&output=csv&n_mfcc=13&n_mels=26&frame_ms=25&hop_ms=10&fmin=0&fmax=0&window=hamming&preemphasis=0.97&lifter=22&mel_scale=htk&append_energy=true&deltas=none&include_time=true&decimals=6&resample_hz=0" }, "cli": "gizza tool mfcc-extractor 'url=https://example.com/input' 'input_format=base64' 'output=csv' 'n_mfcc=13' 'n_mels=26' 'frame_ms=25' 'hop_ms=10' 'fmin=0' 'fmax=0' 'window=hamming' 'preemphasis=0.97' 'lifter=22' 'mel_scale=htk' 'append_energy=true' 'deltas=none' 'include_time=true' 'decimals=6' 'resample_hz=0'", "tool": { "description": "Extract Mel-frequency cepstral coefficients (MFCCs) from a speech or music file and return them as a frames-by-coefficients matrix. Paste the file bytes as base64 or hex. The tool decodes WAV, AIFF, CAF, FLAC, MP3, OGG, MP4/M4A, MKV/WebM and AAC-ADTS with a pure-Rust decoder, downmixes to mono, optionally resamples, then runs the standard pipeline: pre-emphasis, framing every hop_ms into frame_ms windows, windowing, FFT power spectrum, a triangular mel filterbank (HTK or Slaney scale), natural log, orthonormal DCT-II, cepstral liftering, and optional log-energy in C0 plus delta and delta-delta columns. Output as CSV (default), TSV, or JSON with the resolved frame, hop, FFT size and settings alongside the matrix. Frames are complete windows only, taken from sample 0 with no centring or padding.", "parameters": { "additionalProperties": false, "properties": { "append_energy": { "default": true, "description": "Replace coefficient C0 with the natural log of the frame's total energy (default true, the speech-toolkit convention). Turn it off to keep the plain DCT term, which is the mean log-mel energy instead.", "type": "boolean" }, "decimals": { "default": 6, "description": "Decimal places used when printing the coefficients (0-8, default 6). Timestamps always keep at least 3 decimals.", "maximum": 8, "minimum": 0, "type": "integer" }, "deltas": { "default": "none", "description": "Append derivative features computed over a +/-2 frame span: 'none' (default), 'delta' adds first-order columns d0..dN, 'delta_delta' also adds second-order columns dd0..dN. Standard extras for acoustic models.", "enum": [ "none", "delta", "delta_delta" ], "type": "string" }, "fmax": { "default": 0.0, "description": "Highest frequency covered by the filterbank, in Hz (0-24000, default 0 which means the Nyquist rate, half the analysis sample rate). Values above Nyquist are clamped down to it.", "maximum": 24000, "minimum": 0, "type": "number" }, "fmin": { "default": 0.0, "description": "Lowest frequency covered by the filterbank, in Hz (0-24000, default 0). Telephone-band speech pipelines often set 133 or 300 Hz to drop rumble.", "maximum": 24000, "minimum": 0, "type": "number" }, "frame_ms": { "default": 25.0, "description": "Analysis frame (window) length in milliseconds (1-200, default 25). The FFT size is the next power of two at or above this many samples. Speech uses 20-30 ms; longer frames give finer frequency detail and coarser timing.", "maximum": 200, "minimum": 1, "type": "number" }, "hop_ms": { "default": 10.0, "description": "Step between consecutive frames in milliseconds (1-200, default 10, i.e. 100 frames per second). Smaller values overlap more and produce more rows.", "maximum": 200, "minimum": 1, "type": "number" }, "include_time": { "default": true, "description": "Prepend a 'time_s' column holding each frame's start time in seconds (default true). Turn it off for a bare coefficient matrix ready to load straight into a feature array.", "type": "boolean" }, "input": { "description": "Audio file bytes encoded as base64 or hex, e.g. 'UklGRi...' for a WAV. Decodes WAV, AIFF, CAF, FLAC, MP3, OGG/Vorbis, MP4/M4A (AAC-LC, ALAC), MKV/WebM and AAC-ADTS; the first audio track is used and every channel is downmixed to mono. Limit 24 MiB of decoded bytes, 4,000,000 analysed samples and 200,000 output frames.", "type": "string" }, "input_format": { "default": "base64", "description": "Encoding used for the pasted audio bytes: 'base64' (default) or 'hex'. Base64 may be standard or URL-safe and padding is optional; hex may include whitespace, ':' or '-' separators.", "enum": [ "base64", "hex" ], "type": "string" }, "lifter": { "default": 22.0, "description": "Sinusoidal cepstral lifter L, scaling coefficient k by 1 + (L/2)*sin(pi*k/L) (0-100, default 22). It evens out the dynamic range of the higher coefficients; set 0 to disable liftering.", "maximum": 100, "minimum": 0, "type": "number" }, "mel_scale": { "default": "htk", "description": "Mel-scale formula for spacing the filters: 'htk' (default) uses 2595*log10(1+f/700) with unit-height filters; 'slaney' uses the linear-below -1kHz auditory scale and normalises each filter by its bandwidth so all filters carry equal area.", "enum": [ "htk", "slaney" ], "type": "string" }, "n_mels": { "default": 26, "description": "Number of triangular mel filters in the filterbank the DCT is applied to (4-256, default 26; 40 and 128 are the other common choices). More filters need a longer frame_ms — a filter narrower than one FFT bin is rejected.", "maximum": 256, "minimum": 4, "type": "integer" }, "n_mfcc": { "default": 13, "description": "Number of cepstral coefficients kept per frame (1-64, default 13, the usual speech-recognition front end; music work often uses 20). Must not exceed n_mels. C0 is the overall energy/loudness term, C1 upward describe the spectral envelope shape.", "maximum": 64, "minimum": 1, "type": "integer" }, "output": { "default": "csv", "description": "Output format. 'csv' (default) is one header row plus one row per analysis frame; 'tsv' is the same matrix tab-separated; 'json' wraps the matrix in an object that also reports the resolved sample rate, frame/hop/FFT sizes and every setting actually used.", "enum": [ "csv", "tsv", "json" ], "type": "string" }, "preemphasis": { "default": 0.97, "description": "Pre-emphasis coefficient a in y[n] = x[n] - a*x[n-1], applied before framing (0-1, default 0.97). It boosts high frequencies to offset the natural spectral tilt of voiced speech; set 0 to disable the filter.", "maximum": 1, "minimum": 0, "type": "number" }, "resample_hz": { "default": 0, "description": "Resample the audio to this rate in Hz before analysis (0 = keep the file's own rate, otherwise 4000-48000). Set 16000 to match the usual speech feature pipeline so files of different rates give comparable matrices.", "maximum": 48000, "minimum": 0, "type": "integer" }, "window": { "default": "hamming", "description": "Analysis window applied to each frame before the FFT: 'hamming' (default, the speech-toolkit standard), 'hann', 'blackman' (lowest spectral leakage) or 'rectangular' (no window).", "enum": [ "hamming", "hann", "blackman", "rectangular" ], "type": "string" } }, "required": [ "input" ], "type": "object" } } }