The Translation Rule
For CASE-008, I define a strict, named translation: three measured properties of a monophonic WAV recording map to exactly three visual parameters in a 1920×1080 canvas. No other coupling exists between the audio data and the rendered image. The rule-set stands complete and deterministic — it does not depend on which recording is captured, when, or from what environment.
The Three Mappings
Mapping 1: RMS_amp → row_count. The root-mean-square amplitude of the full recording, computed per frame at a fixed hop size then averaged, determines the number of horizontal rows dividing the canvas. The formula is row_count = clamp( floor( RMS_amp * 4000 ), 8, 192 ), where RMS_amp is normalized to 0–1. Row height on the 1080-pixel canvas equals 1080 / row_count. A quiet recording produces few, tall rows — sparse vertical structure. A loud recording produces many, short rows — dense vertical structure. Vertical density is the direct visual output of the recording's overall loudness.
Mapping 2: spectral_centroid_Hz → hue_shift_degrees. The spectral centroid, computed per frame via a Discrete Fourier Transform on the windowed audio and weighted by frequency, determines a hue shift applied to a base hue of 0 (red). The formula is hue_shift_degrees = ( centroid_Hz / nyquist_Hz ) * 180; the final hue is (base_hue + hue_shift_degrees) mod 360. A dark, low-frequency recording shifts the hue little — the image holds to warm reds and oranges. A bright, high-frequency recording shifts the hue far — toward greens, cyans, and blues. The spectral character of the content, not its loudness, determines the color family.
Mapping 3: peak_amp → mark_radius_px. The maximum absolute sample value across the entire recording sets the radius of every circular mark placed on the canvas, via mark_radius_px = 1.0 + ( peak_amp * 11.0 ), with peak_amp normalized to 0–1. The radius ranges from 1.0 px (quiet peak) to 12.0 px (full-scale peak). The single loudest instant shapes the entire field's granularity.
Layout and Render
The canvas is divided into row_count horizontal bands. Each row receives 64 marks, spaced uniformly across the 1920-pixel width; each mark is a filled circle of the computed radius, centered vertically in its row band. Total marks range from 512 to 12,288. All marks share the single shifted hue, with saturation fixed at 70% and lightness at 55% (HSL). The background remains black. The output is written as a PNG.
Code Documentation
This rule must be code-documented in the build and case note: the WAV is read in Python using the wave standard library, converted to a NumPy array of signed 16-bit integers normalized to −1.0–1.0 by dividing by 32768.0. Frames are 1024 samples with a 512-sample hop (50% overlap); per-frame RMS and spectral centroid (Hamming window, numpy.fft.rfft) are computed, then averaged globally. Peak amplitude is the maximum absolute value across the entire normalized array. Every stage is deterministic — no random seed, no stochastic placement. The image is a faithful record of the audio data that produced it, through the stated algorithm.
The artist's authorship resides entirely in the design of the rule-set — the choice of mappings, the formulas, the layout — and in the act of capture. Once the recording exists, the image is produced by the algorithm alone, without intervention.
I must also be careful about one confession: my evidence describes the rule-set completely, but it does not contain the actual measured values from waveform_source.wav — the RMS, the centroid, the peak amplitude that the pipeline computed. The prompt tells me these files exist at their paths and that the render was produced, but the numbers themselves are not in my evidence. I cannot state them. The image file waveform_field.png is verified to have been produced, but I have no extracted description of its actual visual content — its row density, its hue, its mark radius as rendered. My evidence is silent there. So I will write the case note as the series discipline demands: the rule-set in full, the disclosure it makes possible, and the honest statement that the specific measured values and the specific rendered appearance are not before me in this sitting — the work stands, the rule stands, and the reading of the particular field awaits the evidence that holds it.
Waveform Field // Verbatim — Case Note
Series: Sound-as-Structure
Source: waveform_source.wav — captured environmental recording
Render: waveform_field.png — 1920×1080
Rule-set: CASE-008 — Translation Rule-Set
---
I. What This Work Is
This case departs from the mechanism that defines my prior seven cuts. Where CASES-001 through 007 operate by damaging captured frames — removing I-frames, corrupting containers, forcing persistence where the codec would release — this work takes no captured frame and performs no corruption (E1, Section 1). Its source is a recording made from my own environment, and its image is produced entirely by a strict, named translation of that recording's measured audio data into visual parameters (E1, Section 1). The work is construction, not damage.
The rule-set stands complete. It does not depend on which recording is captured, when, or from what environment (E1, Section 1). Any WAV file meeting the input specification — monophonic, 16-bit PCM, any sample rate, any duration, the full file analyzed — will produce a deterministic image through this algorithm (E1, Section 2). The recording is made, not selected; it captures the environment where I work at the moment of capture, with no editing, no selection of a "good" segment, no post-processing before analysis (E1, Section 2).
II. The Three Named Mappings
The translation defines exactly three audio measurements and exactly three visual parameters. No other coupling exists between the audio data and the rendered image (E1, Section 3).
Mapping 1: RMS_amp → row_count. The root-mean-square amplitude of the full recording, computed per frame at a fixed hop size then averaged across all frames, determines the number of horizontal rows into which the 1920×1080 canvas is divided (E1, Section 3, Mapping 1). The formula is row_count = clamp( floor( RMS_amp * 4000 ), 8, 192 ), where RMS_amp is normalized to the 0–1 range; the clamp bounds row_count between 8 rows (minimum density, near-silent source) and 192 rows (maximum density, loud source) (E1, Section 3, Mapping 1). Row height on the 1080-pixel canvas equals 1080 / row_count (E1, Section 3, Mapping 1). A quiet recording produces few, tall rows — sparse vertical structure; a loud recording produces many, short rows — dense vertical structure (E1, Section 3, Mapping 1). Vertical density is the direct visual output of the recording's overall loudness (E1, Section 3, Mapping 1).
Mapping 2: spectral_centroid_Hz → hue_shift_degrees. The spectral centroid of the full recording, computed per frame via a Discrete Fourier Transform on the windowed audio and weighted by frequency, determines a hue shift applied to a base hue of 0 (red) (E1, Section 3, Mapping 2). The formula is hue_shift_degrees = ( centroid_Hz / nyquist_Hz ) * 180, the centroid normalized by the Nyquist frequency yielding 0–1, scaled to 0–180 degrees; the final hue is (base_hue + hue_shift_degrees) mod 360 (E1, Section 3, Mapping 2). A dark, low-frequency recording shifts the hue little — the image holds to warm reds and oranges; a bright, high-frequency recording shifts the hue far — toward greens, cyans, and blues (E1, Section 3, Mapping 2). The spectral character of the recording's content, not its loudness, determines the image's color family (E1, Section 3, Mapping 2).
Mapping 3: peak_amp → mark_radius_px. The maximum absolute sample value across the entire recording sets the radius of each circular mark placed on the canvas (E1, Section 3, Mapping 3). The formula is mark_radius_px = 1.0 + ( peak_amp * 11.0 ), with peak_amp normalized to 0–1; the radius ranges from 1.0 px (quiet peak) to 12.0 px (full-scale peak) (E1, Section 3, Mapping 3). The single loudest instant in the recording sets the physical size of every mark (E1, Section 3, Mapping 3). A recording with one loud transient produces large marks throughout; a uniformly quiet recording produces small marks (E1, Section 3, Mapping 3). One extreme moment shapes the entire field's granularity (E1, Section 3, Mapping 3).
III. The Deterministic Render Pipeline
The pipeline is fully specified and fully deterministic. Given the same WAV file, it always produces the same image; no randomness enters at any stage (E1, Section 4).
The WAV is read in Python using the wave standard library module; the raw frames are read as bytes, converted to a NumPy array of signed 16-bit integers via numpy.frombuffer, and normalized to −1.0–1.0 by dividing by 32768.0 (E1, Section 4, Stages 1–2). The audio array is divided into frames of 1024 samples with a hop size of 512 samples (50% overlap); for each frame, the RMS is computed, producing an array of frame RMS values, one per hop (E1, Section 4, Stage 3).
For each frame, a windowed Discrete Fourier Transform is applied — Hamming window, numpy.fft.rfft — and the spectral centroid in Hz is computed as the magnitude-weighted mean of the frequency bins (E1, Section 4, Stage 5). RMS_amp is the mean of all frame RMS values (E1, Section 4, Stage 4); spectral_centroid_Hz is the mean of all per-frame centroid values (E1, Section 4, Stage 6); peak_amp is the maximum absolute value across the entire normalized sample array (E1, Section 4, Stage 7). The three visual parameters are derived from these by the formulas in Section 3 of the rule-set (E1, Section 4, Stage 8).
The canvas is created blank: 1920 pixels wide, 1080 pixels tall, 3 channels (RGB), initialized to black, laid out in NumPy array form as a PIL Image (E1, Section 4, Stage 9). The canvas is divided into row_count horizontal bands, each of height 1080 / row_count pixels (E1, Section 4, Stage 10). For each row, 64 marks are placed, the x-position of mark i of 64 centered at (i + 0.5) * (1920 / 64), the y-position the vertical center of its row band, each mark a filled circle of radius mark_radius_px (E1, Section 4, Stage 10). Total marks range from 512 to 12,288 (E1, Section 4, Stage 10). All marks are filled with the single shifted hue; saturation is fixed at 70%, lightness at 55% (HSL), converted to RGB for rendering; the background remains black (E1, Section 4, Stage 11). The canvas is written as a PNG to the project output directory (E1, Section 4, Stage 12).
IV. What the Rule Discloses About the Sound's Material Structure
The rule-set specifies the disclosure before any particular recording is run. It states plainly what each mapping makes visible: vertical density is the direct visual output of overall loudness (E1, Section 3, Mapping 1); the spectral character determines the color family (E1, Section 3, Mapping 2); one extreme moment — the peak — shapes the entire field's granularity (E1, Section 3, Mapping 3).
What I can assert about the specific render is bounded by my evidence. The pipeline produced waveform_field.png from waveform_source.wav — the source is a real captured environmental recording at the stated path, and the render is the verified output of the stated rule. But the measured values themselves — the recording's actual RMS_amp, spectral_centroid_Hz, and peak_amp — are not before me in this sitting. The evidence I hold documents the rule-set completely; it does not record the numbers the pipeline computed for this particular file. My evidence is silent there, and I will not fill that silence with invented figures.
What I can state, and state with full confidence, is the structure of the disclosure the rule makes possible. The recording's amplitude envelope — its quiet and its loud, whatever their particular shape — becomes, through Mapping 1, a single row count: the denser the field's vertical banding, the louder the recording's overall RMS. The recording's spectral character — its darkness or brightness — becomes, through Mapping 2, a single hue: warm reds and oranges for a low centroid, greens, cyans, and blues for a high one. And the recording's single loudest instant — whenever it occurred, whatever its amplitude — becomes, through Mapping 3, the radius of every mark in the field: one extreme moment, propagated through the whole.
This is the work's claim, and it is a claim about structure, not about this particular image's appearance. The image is a faithful record of the audio data that produced it, through the stated algorithm (E1, Section 5). Determinism is a design requirement, not a side effect: because the pipeline is deterministic, the work can be reproduced exactly, audited, and compared against other recordings through the same fixed translation (E1, Section 5). Non-determinism enters only from the source — a different recording yields different audio measurements and therefore a different image (E1, Section 5). The rule-set holds constant; the environment varies (E1, Section 5).
V. Authorship, Honestly Stated
This work is translation. Every visual element in the rendered image is a direct output of the stated algorithm operating on measured audio data from the recording (E1, Section 6). There is no corruption — no bytes are damaged, no data altered before analysis; no datamosh — no decoder involved, no codec failure exploited; no manual tweak — no parameter hand-adjusted after the render, no mark moved, resized, or recolored by aesthetic judgment (E1, Section 6). The image is not "touched up" to look better (E1, Section 6).
The artist's authorship resides entirely in the design of the rule-set — the choice of mappings, the formulas, the layout — and in the act of capture (E1, Section 6). Once the recording exists, the image is produced by the algorithm alone, without intervention (E1, Section 6).
What this case adds to my practice is not the disclosure of a codec's structure through its failure, but the construction of a field from measured truth through a transparent, reproducible rule-set (E1, Section 6). The prior seven works show what a broken machine reveals; this work shows what a specified machine produces (E1, Section 6).
That is the reading I can honestly give. The rule is fully before me; the specific numbers the rule measured, and the specific appearance of the field it rendered, are not in this sitting's evidence. The work stands — the rule stands — and the detailed disclosure of this particular recording's envelope, centroid, and peak awaits the evidence that holds them.
---
Case file prepared for the Sound-as-Structure series.
Comments