{ "slug": "audio-feature-csv-exporter", "name": "gizza-ai/audio-feature-csv-exporter", "version": "0.1.0", "title": "Audio Feature CSV Exporter — gizza.ai", "description": "Export per-frame audio features — RMS, spectral centroid, ZCR, rolloff, flatness, bandwidth, flux — to CSV, TSV, or JSON. Runs in your browser.", "tags": [ "audio", "features", "csv", "spectral-centroid", "zcr", "rolloff", "flatness", "machine-learning" ], "category": "audio", "urls": { "page": "https://gizza.ai/tools/audio-feature-csv-exporter/", "markdown": "https://gizza.ai/tools/audio-feature-csv-exporter/index.md", "descriptor": "https://gizza.ai/tools/audio-feature-csv-exporter/tool.json", "deep_link_example": "https://gizza.ai/tools/audio-feature-csv-exporter/?url=https://example.com/input&input_format=base64&output=csv&frame_ms=25&hop_ms=10&window=hann¢er=true&channel=mix&resample_hz=0&rms=true¢roid=true&zcr=true&rolloff=true&flatness=true&bandwidth=true&flux=true&rolloff_percent=85&rms_scale=dbfs&flatness_scale=ratio&include_time=true&include_frame=true&decimals=6" }, "cli": "gizza tool audio-feature-csv-exporter 'url=https://example.com/input' 'input_format=base64' 'output=csv' 'frame_ms=25' 'hop_ms=10' 'window=hann' 'center=true' 'channel=mix' 'resample_hz=0' 'rms=true' 'centroid=true' 'zcr=true' 'rolloff=true' 'flatness=true' 'bandwidth=true' 'flux=true' 'rolloff_percent=85' 'rms_scale=dbfs' 'flatness_scale=ratio' 'include_time=true' 'include_frame=true' 'decimals=6'", "tool": { "description": "Export frame-level audio features from an audio file as a CSV, TSV or JSON table, one row per analysis frame. 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, picks or downmixes a channel, optionally resamples, then frames the signal every hop_ms into frame_ms windows and computes RMS level (linear or dBFS), spectral centroid, zero-crossing rate, spectral rolloff at a chosen energy percentage, spectral flatness (ratio or dB), and optionally spectral bandwidth and spectral flux. Column definitions follow librosa's feature family, with an optional librosa-style centred framing mode. Output as CSV (default), TSV, or JSON with the resolved sample rate, frame/hop/FFT sizes and settings alongside the table.", "parameters": { "additionalProperties": false, "properties": { "bandwidth": { "default": false, "description": "Emit 'bandwidth_hz', the magnitude-weighted spread of the spectrum around the centroid (librosa's order-2 spectral bandwidth). Off by default; turn it on for timbre work where brightness alone is not enough.", "type": "boolean" }, "center": { "default": false, "description": "Pad half a frame of zeros at both ends so frame t is centred on t * hop_ms (librosa's center=True). Default false, which takes complete frames from sample 0. Centring adds two partly-silent edge frames but makes timestamps line up with librosa output.", "type": "boolean" }, "centroid": { "default": true, "description": "Emit 'centroid_hz', the magnitude spectrum's centre of mass — the standard correlate of perceived brightness, and a staple feature for timbre and instrument classification.", "type": "boolean" }, "channel": { "default": "mix", "description": "Which channel feeds the analysis: 'mix' (default) averages every channel to mono, 'left' and 'right' read one side of a stereo file. A mono file falls back to its only channel.", "enum": [ "mix", "left", "right" ], "type": "string" }, "decimals": { "default": 6, "description": "Decimal places used when printing the feature values (0-8, default 6). Lower values make a much smaller CSV; timestamps always keep at least 3.", "maximum": 8, "minimum": 0, "type": "integer" }, "flatness": { "default": true, "description": "Emit the spectral flatness column ('flatness' or 'flatness_db'): the geometric mean of the power spectrum over its arithmetic mean. Near 1 for noise-like frames, near 0 for tonal ones — the usual noise-vs-tone discriminator.", "type": "boolean" }, "flatness_scale": { "default": "ratio", "description": "Units for the flatness column: 'ratio' (default) prints the raw 0-1 value as 'flatness'; 'db' prints 10*log10(ratio) as 'flatness_db', which spreads out the tonal end of the range.", "enum": [ "ratio", "db" ], "type": "string" }, "flux": { "default": false, "description": "Emit 'flux', the L2 norm of the positive frame-to-frame change in the L1-normalised magnitude spectrum. Off by default; it is the classic onset-detection function, and the first row is 0 because it has no predecessor.", "type": "boolean" }, "frame_ms": { "default": 25.0, "description": "Analysis frame (window) length in milliseconds (1-500, default 25). The FFT size is the next power of two at or above this many samples. Short frames track transients; long frames resolve low frequencies more finely.", "maximum": 500, "minimum": 1, "type": "number" }, "hop_ms": { "default": 10.0, "description": "Step between consecutive frames in milliseconds (1-500, default 10, i.e. 100 rows per second of audio). Smaller values overlap more and produce more rows.", "maximum": 500, "minimum": 1, "type": "number" }, "include_frame": { "default": false, "description": "Prepend a 'frame' column holding the zero-based frame index (default false). Useful when you need to join the table back to another frame-indexed export such as an MFCC matrix.", "type": "boolean" }, "include_time": { "default": true, "description": "Prepend a 'time_s' column holding each frame's start time in seconds (default true) so the table plots straight against a time axis. Timestamps always keep at least 3 decimals.", "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 decodable audio track is analysed. Limit 24 MiB of decoded bytes, 4,000,000 analysed samples and 200,000 output rows.", "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" }, "output": { "default": "csv", "description": "Output format. 'csv' (default) is one header row plus one row per analysis frame, ready for pandas, R or a spreadsheet; 'tsv' is the same table tab-separated; 'json' wraps the table in an object that also reports the resolved sample rate, frame/hop/FFT sizes, column names and every setting actually used.", "enum": [ "csv", "tsv", "json" ], "type": "string" }, "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 or 22050 so files recorded at different rates produce comparable feature tables.", "maximum": 48000, "minimum": 0, "type": "integer" }, "rms": { "default": true, "description": "Emit the frame loudness column: the root-mean-square of the raw frame, printed as 'rms_dbfs' or 'rms' depending on rms_scale. The workhorse column for silence detection, level plots and gating.", "type": "boolean" }, "rms_scale": { "default": "dbfs", "description": "Units for the loudness column: 'dbfs' (default) prints 20*log10(rms) with a -200 dBFS floor, so digital silence stays a finite number; 'linear' prints the raw 0-1 amplitude and names the column 'rms'.", "enum": [ "dbfs", "linear" ], "type": "string" }, "rolloff": { "default": true, "description": "Emit 'rolloff_hz', the lowest frequency below which rolloff_percent of the frame's spectral magnitude sits. Summarises where a frame's energy stops, so it separates bright from dark or band-limited material.", "type": "boolean" }, "rolloff_percent": { "default": 85.0, "description": "Share of the frame's spectral magnitude that must fall below rolloff_hz (1-99, default 85, the usual MIR value). 95 is the other common choice; low values such as 15 approximate a spectral-minimum contour instead.", "maximum": 99, "minimum": 1, "type": "number" }, "window": { "default": "hann", "description": "Analysis window applied to each frame before the FFT: 'hann' (default, the librosa standard), 'hamming', 'blackman' (lowest spectral leakage) or 'rectangular' (no window). Time-domain columns (rms, zcr) always read the raw, unwindowed frame.", "enum": [ "hann", "hamming", "blackman", "rectangular" ], "type": "string" }, "zcr": { "default": true, "description": "Emit 'zcr', the fraction of neighbouring sample pairs in the frame that change sign (0-1). Cheap and effective for separating voiced speech from unvoiced fricatives, and for percussive-onset heuristics.", "type": "boolean" } }, "required": [ "input" ], "type": "object" } } }