diff --git a/Cargo.toml b/Cargo.toml
index b765d36..465bafd 100644
--- a/Cargo.toml
+++ b/Cargo.toml
@@ -15,6 +15,10 @@ path = "src/main.rs"
 name = "test_net"
 path = "src/bin/test_net.rs"
 
+[[bin]]
+name = "specview"
+path = "src/bin/specview.rs"
+
 [dependencies]
 anyhow = "1.0.102"
 async-trait = "0.1.89"
diff --git a/src/bin/specview.rs b/src/bin/specview.rs
new file mode 100644
index 0000000..685661f
--- /dev/null
+++ b/src/bin/specview.rs
@@ -0,0 +1,258 @@
+//! `specview` — a developer CLI for the `dsp` toolkit: conjure test signals,
+//! draw spectrograms in the terminal, and measure echo-cancellation quality
+//! (ERLE) on WAV files. Not shipped in the GUI; a measurement aid.
+//!
+//! Usage:
+//!   specview show  <file.wav> [--width N] [--height N] [--max-hz HZ] [--floor DB] [--ascii]
+//!   specview gen   <kind> -o out.wav [--secs S] [--amp A] [--freq F] [--f0 F] [--f1 F] [--seed N]
+//!         kinds: sine | sweep | white | pink | impulse | silence
+//!   specview erle  <before.wav> <after.wav> [--show]
+//!
+//! Examples:
+//!   specview gen sweep -o sweep.wav --secs 3 --f0 100 --f1 12000
+//!   specview show ~/peerspeak-recordings/2026-06-14/peerspeak-*.wav
+//!   specview erle echo_only.wav after_aec.wav --show
+
+use std::collections::HashMap;
+use std::path::Path;
+use std::process::ExitCode;
+
+use peerspeak::dsp::{generators, metrics, render, stft, wav};
+
+const SAMPLE_RATE: u32 = 48_000;
+
+fn main() -> ExitCode {
+    let args: Vec<String> = std::env::args().skip(1).collect();
+    let Some(cmd) = args.first() else {
+        eprintln!("{}", USAGE);
+        return ExitCode::FAILURE;
+    };
+
+    let rest = &args[1..];
+    let result = match cmd.as_str() {
+        "show" => cmd_show(rest),
+        "gen" => cmd_gen(rest),
+        "erle" => cmd_erle(rest),
+        "-h" | "--help" | "help" => {
+            println!("{}", USAGE);
+            Ok(())
+        }
+        other => Err(format!("unknown command {other:?}\n\n{USAGE}")),
+    };
+
+    match result {
+        Ok(()) => ExitCode::SUCCESS,
+        Err(e) => {
+            eprintln!("error: {e}");
+            ExitCode::FAILURE
+        }
+    }
+}
+
+const USAGE: &str = "specview — spectrogram + AEC measurement tool\n\n\
+  specview show <file.wav> [--width N] [--height N] [--max-hz HZ] [--floor DB] [--ascii]\n\
+  specview gen  <kind> -o out.wav [--secs S] [--amp A] [--freq F] [--f0 F] [--f1 F] [--seed N]\n\
+                kinds: sine sweep white pink impulse silence\n\
+  specview erle <before.wav> <after.wav> [--show]";
+
+/// Renders a WAV as a spectrogram plus a numeric summary.
+fn cmd_show(args: &[String]) -> Result<(), String> {
+    let (positional, flags) = parse_args(args);
+    let path = positional.first().ok_or("show needs a <file.wav>")?;
+    let data = wav::read(Path::new(path))?;
+
+    let opts = render::RenderOpts {
+        width: flags.usize_or("width", 100),
+        height: flags.usize_or("height", 30),
+        max_hz: flags.f32_or("max-hz", 12_000.0),
+        floor_db: flags.f32_or("floor", -80.0),
+        ascii: flags.present("ascii"),
+    };
+    let spec = stft::analyze(&data.samples, data.sample_rate, 2048, 512);
+    print!("{}", render::render(&spec, &opts));
+    print_summary(path, &data.samples, &spec);
+    Ok(())
+}
+
+/// Generates a test signal and writes it to a WAV file.
+fn cmd_gen(args: &[String]) -> Result<(), String> {
+    let (positional, flags) = parse_args(args);
+    let kind = positional.first().ok_or("gen needs a <kind>")?;
+    let out = flags
+        .get("o")
+        .or_else(|| flags.get("out"))
+        .ok_or("gen needs -o <out.wav>")?;
+
+    let secs = flags.f32_or("secs", 2.0);
+    let amp = flags.f32_or("amp", 0.8);
+    let len = (secs * SAMPLE_RATE as f32) as usize;
+    let seed = flags.u64_or("seed", 1);
+
+    let samples = match kind.as_str() {
+        "sine" => generators::sine(flags.f32_or("freq", 1000.0), amp, len, SAMPLE_RATE),
+        "sweep" => generators::log_sweep(
+            flags.f32_or("f0", 100.0),
+            flags.f32_or("f1", 12_000.0),
+            amp,
+            len,
+            SAMPLE_RATE,
+        ),
+        "white" => generators::white_noise(amp, len, seed),
+        "pink" => generators::pink_noise(amp, len, seed),
+        "impulse" => generators::impulse(amp, len),
+        "silence" => generators::silence(len),
+        other => return Err(format!("unknown kind {other:?} (sine sweep white pink impulse silence)")),
+    };
+
+    wav::write(Path::new(out), &samples, SAMPLE_RATE)?;
+    println!(
+        "wrote {} — {} ({:.2}s, {} samples, rms {:.1} dBFS)",
+        out,
+        kind,
+        secs,
+        samples.len(),
+        metrics::dbfs(metrics::rms(&samples)),
+    );
+    Ok(())
+}
+
+/// Measures ERLE between a "before" (echo present) and "after" (post-AEC) file —
+/// broadband and per voice band — so you can see how much echo was removed and
+/// in which frequency range any residual lives.
+fn cmd_erle(args: &[String]) -> Result<(), String> {
+    let (positional, flags) = parse_args(args);
+    let before = positional.first().ok_or("erle needs <before.wav> <after.wav>")?;
+    let after = positional.get(1).ok_or("erle needs <before.wav> <after.wav>")?;
+
+    let b = wav::read(Path::new(before))?;
+    let a = wav::read(Path::new(after))?;
+    // Compare over the overlapping length.
+    let n = b.samples.len().min(a.samples.len());
+    let bs = &b.samples[..n];
+    let as_ = &a.samples[..n];
+
+    let broadband = metrics::erle(bs, as_);
+    println!("ERLE (broadband): {broadband:+.1} dB   (higher = more echo removed)");
+    println!(
+        "  before: {:.1} dBFS rms   after: {:.1} dBFS rms",
+        metrics::dbfs(metrics::rms(bs)),
+        metrics::dbfs(metrics::rms(as_)),
+    );
+
+    // Per-band ERLE via averaged STFT magnitude energy.
+    let sb = stft::analyze(bs, b.sample_rate, 2048, 512);
+    let sa = stft::analyze(as_, a.sample_rate, 2048, 512);
+    let avg_b = average_frame(&sb);
+    let avg_a = average_frame(&sa);
+    println!("  per band:");
+    for band in metrics::VOICE_BANDS {
+        let eb = metrics::band_energy(&avg_b, |bin| sb.bin_hz(bin), band.low_hz, band.high_hz);
+        let ea = metrics::band_energy(&avg_a, |bin| sa.bin_hz(bin), band.low_hz, band.high_hz);
+        let db = if eb > 1e-12 && ea > 1e-12 {
+            10.0 * (eb / ea).log10()
+        } else if eb > 1e-12 {
+            120.0
+        } else {
+            0.0
+        };
+        println!("    {:<18} {:+6.1} dB", band.label, db);
+    }
+
+    if flags.present("show") {
+        println!("\n--- before ---");
+        print!("{}", render::render(&sb, &render::RenderOpts::default()));
+        println!("\n--- after ---");
+        print!("{}", render::render(&sa, &render::RenderOpts::default()));
+    }
+    Ok(())
+}
+
+/// Prints RMS/peak/dBFS plus the per-band energy distribution of a signal.
+fn print_summary(path: &str, samples: &[f32], spec: &stft::Spectrogram) {
+    println!(
+        "\n{}\n  duration {:.2}s   rms {:.1} dBFS   peak {:.1} dBFS",
+        path,
+        samples.len() as f32 / spec.sample_rate as f32,
+        metrics::dbfs(metrics::rms(samples)),
+        metrics::dbfs(metrics::peak(samples)),
+    );
+    let avg = average_frame(spec);
+    let total: f32 = avg.iter().map(|m| m * m).sum::<f32>().max(1e-12);
+    println!("  energy by band:");
+    for band in metrics::VOICE_BANDS {
+        let e = metrics::band_energy(&avg, |bin| spec.bin_hz(bin), band.low_hz, band.high_hz);
+        let pct = 100.0 * e / total;
+        let bar = "█".repeat((pct / 2.5).round() as usize);
+        println!("    {:<18} {:5.1}%  {}", band.label, pct, bar);
+    }
+}
+
+/// Mean magnitude spectrum across all STFT frames (the time-averaged spectrum).
+fn average_frame(spec: &stft::Spectrogram) -> Vec<f32> {
+    let bins = spec.bins();
+    let mut avg = vec![0.0f32; bins];
+    if spec.frames.is_empty() {
+        return avg;
+    }
+    for frame in &spec.frames {
+        for (bin, a) in avg.iter_mut().enumerate() {
+            *a += frame[bin];
+        }
+    }
+    let n = spec.frames.len() as f32;
+    for a in &mut avg {
+        *a /= n;
+    }
+    avg
+}
+
+/// A trivial `--flag value` / `--flag` (boolean) parser. Positional args are
+/// everything not consumed as a flag value. Keeps the binary dependency-free.
+struct Flags {
+    map: HashMap<String, String>,
+    bools: Vec<String>,
+}
+
+impl Flags {
+    fn get(&self, key: &str) -> Option<&String> {
+        self.map.get(key)
+    }
+    fn present(&self, key: &str) -> bool {
+        self.bools.iter().any(|b| b == key) || self.map.contains_key(key)
+    }
+    fn f32_or(&self, key: &str, default: f32) -> f32 {
+        self.map.get(key).and_then(|v| v.parse().ok()).unwrap_or(default)
+    }
+    fn usize_or(&self, key: &str, default: usize) -> usize {
+        self.map.get(key).and_then(|v| v.parse().ok()).unwrap_or(default)
+    }
+    fn u64_or(&self, key: &str, default: u64) -> u64 {
+        self.map.get(key).and_then(|v| v.parse().ok()).unwrap_or(default)
+    }
+}
+
+/// Splits args into positionals and flags. A token like `--width`/`-o` followed
+/// by a non-flag token captures that token as its value; otherwise it's a bool.
+fn parse_args(args: &[String]) -> (Vec<String>, Flags) {
+    let mut positional = Vec::new();
+    let mut map = HashMap::new();
+    let mut bools = Vec::new();
+    let mut i = 0;
+    while i < args.len() {
+        let tok = &args[i];
+        if let Some(key) = tok.strip_prefix("--").or_else(|| tok.strip_prefix('-')) {
+            let next_is_value = args.get(i + 1).is_some_and(|n| !n.starts_with('-'));
+            if next_is_value {
+                map.insert(key.to_string(), args[i + 1].clone());
+                i += 2;
+            } else {
+                bools.push(key.to_string());
+                i += 1;
+            }
+        } else {
+            positional.push(tok.clone());
+            i += 1;
+        }
+    }
+    (positional, Flags { map, bools })
+}
diff --git a/src/dsp/fft.rs b/src/dsp/fft.rs
new file mode 100644
index 0000000..7ec15ff
--- /dev/null
+++ b/src/dsp/fft.rs
@@ -0,0 +1,186 @@
+//! A small, dependency-free radix-2 Cooley-Tukey FFT.
+//!
+//! We hand-roll this rather than pull in `rustfft`/`realfft` because (a) the
+//! whole DSP toolkit is a developer/measurement aid, not a hot real-time path,
+//! and (b) it keeps the supply-chain surface at zero new crates. The transform
+//! is the textbook iterative in-place algorithm (bit-reversal permutation +
+//! log2(N) butterfly stages), computed in `f64` for headroom even though the
+//! audio it analyses is `f32`.
+//!
+//! Only power-of-two lengths are supported; the STFT layer always pads frames
+//! up to a power of two before calling in.
+
+use std::f64::consts::PI;
+
+/// A minimal complex number for the transform. Kept local (rather than pulling a
+/// `num-complex` dependency) since the FFT is the only thing that needs it.
+#[derive(Clone, Copy, Debug, PartialEq)]
+pub struct Complex {
+    pub re: f64,
+    pub im: f64,
+}
+
+impl Complex {
+    pub const ZERO: Complex = Complex { re: 0.0, im: 0.0 };
+
+    pub fn new(re: f64, im: f64) -> Self {
+        Complex { re, im }
+    }
+
+    /// Magnitude `sqrt(re^2 + im^2)`.
+    pub fn magnitude(self) -> f64 {
+        self.re.hypot(self.im)
+    }
+
+    fn add(self, o: Complex) -> Complex {
+        Complex::new(self.re + o.re, self.im + o.im)
+    }
+
+    fn sub(self, o: Complex) -> Complex {
+        Complex::new(self.re - o.re, self.im - o.im)
+    }
+
+    fn mul(self, o: Complex) -> Complex {
+        Complex::new(
+            self.re * o.re - self.im * o.im,
+            self.re * o.im + self.im * o.re,
+        )
+    }
+}
+
+/// In-place forward FFT. `buf.len()` must be a power of two.
+///
+/// Uses the standard sign convention `X[k] = sum_n x[n] * exp(-2πi·kn/N)`.
+pub fn fft(buf: &mut [Complex]) {
+    transform(buf, false);
+}
+
+/// In-place inverse FFT (normalized by `1/N`), the exact inverse of [`fft`].
+pub fn ifft(buf: &mut [Complex]) {
+    transform(buf, true);
+    let n = buf.len() as f64;
+    for c in buf.iter_mut() {
+        c.re /= n;
+        c.im /= n;
+    }
+}
+
+fn transform(buf: &mut [Complex], inverse: bool) {
+    let n = buf.len();
+    assert!(n.is_power_of_two(), "FFT length {n} must be a power of two");
+    if n <= 1 {
+        return;
+    }
+
+    // Bit-reversal permutation: reorder so the iterative butterflies can run
+    // bottom-up in place.
+    let mut j = 0usize;
+    for i in 1..n {
+        let mut bit = n >> 1;
+        while j & bit != 0 {
+            j ^= bit;
+            bit >>= 1;
+        }
+        j ^= bit;
+        if i < j {
+            buf.swap(i, j);
+        }
+    }
+
+    // Butterfly stages: combine length-`len` DFTs from length-`len/2` halves,
+    // doubling `len` each pass.
+    let sign = if inverse { 1.0 } else { -1.0 };
+    let mut len = 2;
+    while len <= n {
+        let ang = sign * 2.0 * PI / len as f64;
+        let wlen = Complex::new(ang.cos(), ang.sin());
+        let mut i = 0;
+        while i < n {
+            let mut w = Complex::new(1.0, 0.0);
+            for k in 0..len / 2 {
+                let u = buf[i + k];
+                let v = buf[i + k + len / 2].mul(w);
+                buf[i + k] = u.add(v);
+                buf[i + k + len / 2] = u.sub(v);
+                w = w.mul(wlen);
+            }
+            i += len;
+        }
+        len <<= 1;
+    }
+}
+
+/// Forward FFT of a real signal, returning the **one-sided** magnitude spectrum:
+/// bins `0..=N/2` (DC through Nyquist), where `N` is the next power of two ≥
+/// `samples.len()`. The input is zero-padded up to `N`.
+///
+/// Magnitudes are raw (un-normalized) linear amplitudes; callers convert to dB
+/// or normalize as needed.
+pub fn real_magnitude_spectrum(samples: &[f32]) -> Vec<f32> {
+    let n = samples.len().next_power_of_two().max(2);
+    let mut buf = vec![Complex::ZERO; n];
+    for (i, &s) in samples.iter().enumerate() {
+        buf[i].re = s as f64;
+    }
+    fft(&mut buf);
+    buf[..=n / 2].iter().map(|c| c.magnitude() as f32).collect()
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    fn approx(a: f64, b: f64, eps: f64) -> bool {
+        (a - b).abs() <= eps
+    }
+
+    #[test]
+    fn impulse_transforms_to_flat_spectrum() {
+        // FFT of a unit impulse at n=0 is all-ones (flat spectrum).
+        let mut buf = vec![Complex::ZERO; 8];
+        buf[0] = Complex::new(1.0, 0.0);
+        fft(&mut buf);
+        for c in &buf {
+            assert!(approx(c.magnitude(), 1.0, 1e-9), "expected flat 1.0, got {c:?}");
+        }
+    }
+
+    #[test]
+    fn single_bin_sine_peaks_in_that_bin() {
+        // A cosine at exactly bin k=2 over N=16 should put all energy in bin 2
+        // (and its mirror N-2). Check the one-sided spectrum peaks at bin 2.
+        let n = 16;
+        let k = 2;
+        let samples: Vec<f32> = (0..n)
+            .map(|i| (2.0 * PI * k as f64 * i as f64 / n as f64).cos() as f32)
+            .collect();
+        let mag = real_magnitude_spectrum(&samples);
+        let peak_bin = mag
+            .iter()
+            .enumerate()
+            .max_by(|a, b| a.1.partial_cmp(b.1).unwrap())
+            .unwrap()
+            .0;
+        assert_eq!(peak_bin, k, "energy should land in bin {k}, got {peak_bin}");
+    }
+
+    #[test]
+    fn ifft_inverts_fft() {
+        let original: Vec<Complex> = (0..32)
+            .map(|i| Complex::new((i as f64 * 0.3).sin(), (i as f64 * 0.1).cos()))
+            .collect();
+        let mut buf = original.clone();
+        fft(&mut buf);
+        ifft(&mut buf);
+        for (a, b) in original.iter().zip(&buf) {
+            assert!(approx(a.re, b.re, 1e-9) && approx(a.im, b.im, 1e-9));
+        }
+    }
+
+    #[test]
+    #[should_panic(expected = "power of two")]
+    fn non_power_of_two_panics() {
+        let mut buf = vec![Complex::ZERO; 6];
+        fft(&mut buf);
+    }
+}
diff --git a/src/dsp/generators.rs b/src/dsp/generators.rs
new file mode 100644
index 0000000..b6fa57b
--- /dev/null
+++ b/src/dsp/generators.rs
@@ -0,0 +1,156 @@
+//! Reproducible test-signal generators — the "conjured noises" the measurement
+//! harness drives AEC/analysis with. Everything is deterministic: the noise
+//! sources take an explicit seed so a run is byte-for-byte repeatable, which is
+//! what makes A/B comparisons (old build vs new, tuning X vs Y) meaningful.
+//!
+//! All generators emit `f32` samples nominally in `[-amp, amp]` at the given
+//! sample rate. Convert to `i16` at the I/O edge.
+
+use std::f64::consts::PI;
+
+/// A tiny deterministic PRNG (xorshift64*) so the noise generators need no
+/// external crate and produce identical output for a given seed across runs.
+struct XorShift64 {
+    state: u64,
+}
+
+impl XorShift64 {
+    fn new(seed: u64) -> Self {
+        // Avoid the all-zero fixed point.
+        XorShift64 {
+            state: seed.wrapping_mul(0x9E37_79B9_7F4A_7C15) | 1,
+        }
+    }
+
+    fn next_u64(&mut self) -> u64 {
+        let mut x = self.state;
+        x ^= x >> 12;
+        x ^= x << 25;
+        x ^= x >> 27;
+        self.state = x;
+        x.wrapping_mul(0x2545_F491_4F6C_DD1D)
+    }
+
+    /// Uniform `f32` in `[-1.0, 1.0)`.
+    fn next_bipolar(&mut self) -> f32 {
+        // Top 24 bits -> [0,1), then map to [-1,1).
+        let u = (self.next_u64() >> 40) as f32 / (1u32 << 24) as f32;
+        u * 2.0 - 1.0
+    }
+}
+
+/// `len` samples of silence.
+pub fn silence(len: usize) -> Vec<f32> {
+    vec![0.0; len]
+}
+
+/// A pure sine tone at `freq` Hz.
+pub fn sine(freq: f32, amp: f32, len: usize, sample_rate: u32) -> Vec<f32> {
+    (0..len)
+        .map(|i| amp * (2.0 * PI * freq as f64 * i as f64 / sample_rate as f64).sin() as f32)
+        .collect()
+}
+
+/// A unit impulse: `amp` at sample 0, silence after. Its flat spectrum makes it
+/// the natural probe for an echo path's impulse response.
+pub fn impulse(amp: f32, len: usize) -> Vec<f32> {
+    let mut v = vec![0.0; len];
+    if len > 0 {
+        v[0] = amp;
+    }
+    v
+}
+
+/// A logarithmic (exponential) sine sweep from `f0` to `f1` Hz over the buffer —
+/// the standard excitation for measuring a system's frequency response, since it
+/// spends equal time per octave.
+pub fn log_sweep(f0: f32, f1: f32, amp: f32, len: usize, sample_rate: u32) -> Vec<f32> {
+    if len == 0 {
+        return Vec::new();
+    }
+    let t_total = len as f64 / sample_rate as f64;
+    let (f0, f1) = (f0 as f64, f1 as f64);
+    let k = (f1 / f0).ln();
+    (0..len)
+        .map(|i| {
+            let t = i as f64 / sample_rate as f64;
+            // Instantaneous-phase integral of an exponential chirp.
+            let phase = 2.0 * PI * f0 * t_total / k * ((k * t / t_total).exp() - 1.0);
+            amp * phase.sin() as f32
+        })
+        .collect()
+}
+
+/// Uniform white noise in `[-amp, amp]`, seeded for reproducibility.
+pub fn white_noise(amp: f32, len: usize, seed: u64) -> Vec<f32> {
+    let mut rng = XorShift64::new(seed);
+    (0..len).map(|_| amp * rng.next_bipolar()).collect()
+}
+
+/// Pink (1/f) noise via the Voss-McCartney algorithm — perceptually flatter than
+/// white noise and a closer stand-in for room/voice spectra. Seeded.
+pub fn pink_noise(amp: f32, len: usize, seed: u64) -> Vec<f32> {
+    let mut rng = XorShift64::new(seed);
+    const ROWS: usize = 16;
+    let mut rows = [0.0f32; ROWS];
+    let mut running = 0.0f32;
+    let mut out = Vec::with_capacity(len);
+    for i in 0..len {
+        // Each row updates half as often as the previous: the row to refresh is
+        // the index of the lowest set bit of the sample counter.
+        let n = i + 1;
+        let row = (n & n.wrapping_neg()).trailing_zeros() as usize;
+        if row < ROWS {
+            running -= rows[row];
+            rows[row] = rng.next_bipolar();
+            running += rows[row];
+        }
+        // Normalize: sum of ROWS unit-bipolar rows spans roughly [-ROWS, ROWS].
+        out.push(amp * running / ROWS as f32);
+    }
+    out
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use crate::dsp::metrics::rms;
+
+    #[test]
+    fn sine_has_expected_length_and_range() {
+        let s = sine(1000.0, 0.5, 480, 48_000);
+        assert_eq!(s.len(), 480);
+        assert!(s.iter().all(|&x| x.abs() <= 0.5001));
+    }
+
+    #[test]
+    fn impulse_is_single_spike() {
+        let s = impulse(1.0, 10);
+        assert_eq!(s[0], 1.0);
+        assert!(s[1..].iter().all(|&x| x == 0.0));
+    }
+
+    #[test]
+    fn white_noise_is_deterministic_for_a_seed() {
+        let a = white_noise(1.0, 256, 42);
+        let b = white_noise(1.0, 256, 42);
+        let c = white_noise(1.0, 256, 43);
+        assert_eq!(a, b, "same seed -> identical output");
+        assert_ne!(a, c, "different seed -> different output");
+    }
+
+    #[test]
+    fn noise_energy_is_nonzero_and_bounded() {
+        let w = white_noise(0.5, 4096, 7);
+        let r = rms(&w);
+        assert!(r > 0.0 && r < 0.5, "rms {r} should be in (0, amp)");
+    }
+
+    #[test]
+    fn log_sweep_spans_the_buffer() {
+        let s = log_sweep(100.0, 8000.0, 1.0, 4800, 48_000);
+        assert_eq!(s.len(), 4800);
+        // Non-trivial energy present.
+        assert!(rms(&s) > 0.1);
+    }
+}
diff --git a/src/dsp/metrics.rs b/src/dsp/metrics.rs
new file mode 100644
index 0000000..8285566
--- /dev/null
+++ b/src/dsp/metrics.rs
@@ -0,0 +1,132 @@
+//! Objective signal metrics — the numbers that turn "sounds better" into a
+//! measurement. The headline one for echo cancellation is **ERLE** (Echo Return
+//! Loss Enhancement): how much echo energy the canceller removed, in dB.
+
+/// Root-mean-square level of a signal (linear amplitude). `0.0` for empty input.
+pub fn rms(samples: &[f32]) -> f32 {
+    if samples.is_empty() {
+        return 0.0;
+    }
+    let sum_sq: f64 = samples.iter().map(|&s| s as f64 * s as f64).sum();
+    (sum_sq / samples.len() as f64).sqrt() as f32
+}
+
+/// Peak absolute amplitude. `0.0` for empty input.
+pub fn peak(samples: &[f32]) -> f32 {
+    samples.iter().fold(0.0, |m, &s| m.max(s.abs()))
+}
+
+/// Converts a linear amplitude (e.g. an RMS or peak value, relative to full-scale
+/// `1.0`) to decibels below full scale. A floor of -120 dBFS is returned for
+/// silence so the result is always finite.
+pub fn dbfs(linear: f32) -> f32 {
+    if linear <= 1e-6 {
+        return -120.0;
+    }
+    20.0 * linear.log10()
+}
+
+/// **Echo Return Loss Enhancement**, in dB: `10·log10(E[echo²] / E[residual²])`.
+///
+/// `echo` is the signal *before* cancellation (the echo the mic picked up);
+/// `residual` is what's *left after* the canceller ran. A larger number is
+/// better — e.g. +30 dB means the canceller removed 99.9% of the echo energy.
+/// Returns a +120 dB ceiling if the residual is effectively silent (perfect
+/// cancellation) and 0.0 if there was no echo energy to begin with.
+pub fn erle(echo: &[f32], residual: &[f32]) -> f32 {
+    let e_echo = mean_square(echo);
+    let e_res = mean_square(residual);
+    if e_echo <= 1e-12 {
+        return 0.0;
+    }
+    if e_res <= 1e-12 {
+        return 120.0;
+    }
+    10.0 * (e_echo / e_res).log10() as f32
+}
+
+/// Mean square (average energy per sample) of a signal. `0.0` for empty input.
+fn mean_square(samples: &[f32]) -> f64 {
+    if samples.is_empty() {
+        return 0.0;
+    }
+    samples.iter().map(|&s| s as f64 * s as f64).sum::<f64>() / samples.len() as f64
+}
+
+/// A frequency band for per-band energy analysis, in Hz.
+pub struct Band {
+    pub label: &'static str,
+    pub low_hz: f32,
+    pub high_hz: f32,
+}
+
+/// Voice-relevant bands for spotting *where* residual echo or noise lives.
+pub const VOICE_BANDS: &[Band] = &[
+    Band { label: "low (80-300)", low_hz: 80.0, high_hz: 300.0 },
+    Band { label: "low-mid (300-1k)", low_hz: 300.0, high_hz: 1000.0 },
+    Band { label: "mid (1k-3k)", low_hz: 1000.0, high_hz: 3000.0 },
+    Band { label: "high-mid (3k-6k)", low_hz: 3000.0, high_hz: 6000.0 },
+    Band { label: "high (6k-12k)", low_hz: 6000.0, high_hz: 12000.0 },
+];
+
+/// Sums the linear magnitude energy within `[low_hz, high_hz)` across a single
+/// STFT magnitude frame. `bin_hz` maps a bin index to its centre frequency.
+pub fn band_energy(frame: &[f32], bin_hz: impl Fn(usize) -> f32, low_hz: f32, high_hz: f32) -> f32 {
+    frame
+        .iter()
+        .enumerate()
+        .filter(|&(bin, _)| {
+            let hz = bin_hz(bin);
+            hz >= low_hz && hz < high_hz
+        })
+        .map(|(_, &m)| m * m)
+        .sum()
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn rms_of_constant_is_that_constant() {
+        assert!((rms(&[0.5; 100]) - 0.5).abs() < 1e-6);
+        assert_eq!(rms(&[]), 0.0);
+    }
+
+    #[test]
+    fn peak_finds_max_magnitude() {
+        assert_eq!(peak(&[0.1, -0.9, 0.3]), 0.9);
+    }
+
+    #[test]
+    fn dbfs_landmarks() {
+        assert!((dbfs(1.0) - 0.0).abs() < 1e-4, "full scale = 0 dBFS");
+        assert!((dbfs(0.5) - -6.0206).abs() < 1e-3, "half = ~-6 dB");
+        assert_eq!(dbfs(0.0), -120.0, "silence floors");
+    }
+
+    #[test]
+    fn erle_halving_energy_is_about_3db() {
+        // residual amplitude = echo/sqrt(2) -> half the energy -> ~3.01 dB.
+        let echo = vec![1.0f32; 1000];
+        let residual = vec![std::f32::consts::FRAC_1_SQRT_2; 1000];
+        let e = erle(&echo, &residual);
+        assert!((e - 3.0103).abs() < 0.01, "expected ~3 dB, got {e}");
+    }
+
+    #[test]
+    fn erle_perfect_cancellation_ceils() {
+        assert_eq!(erle(&[1.0; 10], &[0.0; 10]), 120.0);
+        assert_eq!(erle(&[0.0; 10], &[0.0; 10]), 0.0, "no echo -> 0");
+    }
+
+    #[test]
+    fn band_energy_selects_the_right_bins() {
+        // 5-bin frame, 100 Hz per bin: bins at 0,100,200,300,400 Hz.
+        let frame = [1.0, 2.0, 3.0, 4.0, 5.0];
+        let bin_hz = |b: usize| b as f32 * 100.0;
+        // [150,350) -> bins 200,300 Hz -> 3^2 + 4^2 = 25.
+        let e = band_energy(&frame, bin_hz, 150.0, 350.0);
+        assert!((e - 25.0).abs() < 1e-4, "got {e}");
+    }
+}
diff --git a/src/dsp/mod.rs b/src/dsp/mod.rs
new file mode 100644
index 0000000..96f8ae6
--- /dev/null
+++ b/src/dsp/mod.rs
@@ -0,0 +1,18 @@
+//! Signal-analysis toolkit: FFT, STFT spectrograms, reproducible test-signal
+//! generators, objective metrics (RMS/dBFS/ERLE/per-band energy), a minimal WAV
+//! reader/writer, and a 24-bit-colour terminal spectrogram renderer.
+//!
+//! This is a **measurement/developer aid**, not part of the real-time audio
+//! path. It exists so audio work — especially echo-cancellation tuning — can be
+//! evaluated with reproducible numbers and a visible spectrogram instead of
+//! ear alone. The pure layers (`fft`, `window`, `stft`, `generators`, `metrics`,
+//! `render`) have no I/O; `wav` is the only edge that touches the filesystem.
+//! Driven from the `specview` binary (`src/bin/specview.rs`).
+
+pub mod fft;
+pub mod generators;
+pub mod metrics;
+pub mod render;
+pub mod stft;
+pub mod wav;
+pub mod window;
diff --git a/src/dsp/render.rs b/src/dsp/render.rs
new file mode 100644
index 0000000..161c921
--- /dev/null
+++ b/src/dsp/render.rs
@@ -0,0 +1,260 @@
+//! Renders a [`Spectrogram`] to a string for the terminal.
+//!
+//! The trick that gives real resolution in a text grid: each character cell is
+//! the Unicode upper-half block `▀` with a 24-bit-colour **foreground** (top
+//! pixel) and **background** (bottom pixel), so one row of text shows two
+//! frequency bins. Magnitudes are mapped to dB against the loudest cell and
+//! coloured with a magma-style ramp (dark = quiet, bright = loud). An ASCII
+//! fallback ramp is available for non-truecolor / piped output.
+
+use super::stft::Spectrogram;
+
+/// Rendering options for [`render`].
+pub struct RenderOpts {
+    /// Output width in character columns (time axis).
+    pub width: usize,
+    /// Output height in character rows; vertical resolution is `2×height` bins.
+    pub height: usize,
+    /// Dynamic-range floor in dB below the loudest cell (e.g. `-80.0`).
+    pub floor_db: f32,
+    /// Top of the frequency axis in Hz (clamped to Nyquist).
+    pub max_hz: f32,
+    /// Use a plain-ASCII brightness ramp instead of 24-bit colour blocks.
+    pub ascii: bool,
+}
+
+impl Default for RenderOpts {
+    fn default() -> Self {
+        RenderOpts {
+            width: 100,
+            height: 30,
+            floor_db: -80.0,
+            max_hz: 12_000.0,
+            ascii: false,
+        }
+    }
+}
+
+const RESET: &str = "\x1b[0m";
+/// Brightness ramp (dark→bright) for the ASCII fallback.
+const ASCII_RAMP: &[u8] = b" .:-=+*#%@";
+
+/// Renders `spec` to a multi-line string: a header, the frequency-labelled
+/// spectrogram body, a time axis, and a dB colour legend.
+pub fn render(spec: &Spectrogram, opts: &RenderOpts) -> String {
+    if spec.frames.is_empty() {
+        return "(empty signal — nothing to plot)\n".to_string();
+    }
+
+    let nyquist = spec.sample_rate as f32 / 2.0;
+    let max_hz = opts.max_hz.min(nyquist).max(spec.bin_hz(1));
+    // Number of bins from DC up to max_hz (at least 2 so we can aggregate).
+    let bins_used = (0..spec.bins())
+        .take_while(|&b| spec.bin_hz(b) <= max_hz)
+        .count()
+        .max(2);
+
+    let n_frames = spec.frames.len();
+    let width = opts.width.clamp(1, n_frames);
+    let rows = opts.height.max(1) * 2; // 2 frequency pixels per character row
+
+    // Stage 1: aggregate frames into `width` time columns (mean magnitude/bin).
+    let mut columns: Vec<Vec<f32>> = Vec::with_capacity(width);
+    for c in 0..width {
+        let f_lo = c * n_frames / width;
+        let f_hi = ((c + 1) * n_frames / width).max(f_lo + 1).min(n_frames);
+        let span = (f_hi - f_lo) as f32;
+        let mut acc = vec![0.0f32; bins_used];
+        for f in f_lo..f_hi {
+            let frame = &spec.frames[f];
+            for (bin, a) in acc.iter_mut().enumerate() {
+                *a += frame[bin];
+            }
+        }
+        for a in &mut acc {
+            *a /= span;
+        }
+        columns.push(acc);
+    }
+
+    // Stage 2: aggregate each column's bins into `rows` vertical pixels.
+    // grid[col][pixel], pixel 0 = lowest freq.
+    let mut grid: Vec<Vec<f32>> = Vec::with_capacity(width);
+    let mut max_mag = 1e-9f32;
+    for col in &columns {
+        let mut pixels = vec![0.0f32; rows];
+        for (p, px) in pixels.iter_mut().enumerate() {
+            let b_lo = p * bins_used / rows;
+            let b_hi = ((p + 1) * bins_used / rows).max(b_lo + 1).min(bins_used);
+            let mean = col[b_lo..b_hi].iter().copied().sum::<f32>() / (b_hi - b_lo) as f32;
+            *px = mean;
+            max_mag = max_mag.max(mean);
+        }
+        grid.push(pixels);
+    }
+
+    // Stage 3: draw. Char row r (0 = top) shows pixels (top=high freq, bottom).
+    let mut out = String::new();
+    out.push_str(&format!(
+        "Spectrogram  {:.0} Hz–{:.0} Hz  ·  {:.2}s  ·  {} frames, fft {}, hop {}\n",
+        0.0,
+        max_hz,
+        spec.frame_time(n_frames),
+        n_frames,
+        spec.fft_size,
+        spec.hop,
+    ));
+
+    let label_every = (opts.height.max(1) / 6).max(1);
+    for r in 0..opts.height {
+        let top_pixel = rows - 1 - 2 * r;
+        let bot_pixel = rows.saturating_sub(2 + 2 * r);
+        // Frequency label for this char row (use the top pixel's centre freq).
+        let label = if r % label_every == 0 {
+            let frac = top_pixel as f32 / (rows - 1) as f32;
+            format!("{:>6.0}", frac * max_hz)
+        } else {
+            "      ".to_string()
+        };
+        out.push_str(&label);
+        out.push_str(" \u{2502}"); // " │"
+        for col in &grid {
+            let top = norm_db(col[top_pixel], max_mag, opts.floor_db);
+            let bot = norm_db(col[bot_pixel], max_mag, opts.floor_db);
+            push_cell(&mut out, top, bot, opts.ascii);
+        }
+        if !opts.ascii {
+            out.push_str(RESET);
+        }
+        out.push('\n');
+    }
+
+    // Time axis.
+    out.push_str("       \u{2514}"); // "       └"
+    out.push_str(&"\u{2500}".repeat(width));
+    out.push('\n');
+    out.push_str(&format!(
+        "        0.00s{:>width$.2}s\n",
+        spec.frame_time(n_frames),
+        width = width.saturating_sub(5),
+    ));
+
+    out.push_str(&legend(opts));
+    out
+}
+
+/// Maps a linear magnitude to `[0,1]` over `[floor_db, 0]` relative to `max_mag`.
+fn norm_db(mag: f32, max_mag: f32, floor_db: f32) -> f32 {
+    if mag <= 1e-9 {
+        return 0.0;
+    }
+    let db = 20.0 * (mag / max_mag).log10();
+    ((db - floor_db) / -floor_db).clamp(0.0, 1.0)
+}
+
+/// Appends one character cell (two stacked pixels) to `out`.
+fn push_cell(out: &mut String, top: f32, bot: f32, ascii: bool) {
+    if ascii {
+        // One pixel per cell in ASCII mode (use the brighter of the pair).
+        let v = top.max(bot);
+        let idx = ((v * (ASCII_RAMP.len() - 1) as f32).round() as usize).min(ASCII_RAMP.len() - 1);
+        out.push(ASCII_RAMP[idx] as char);
+    } else {
+        let (tr, tg, tb) = magma(top);
+        let (br, bg, bb) = magma(bot);
+        // fg = top pixel, bg = bottom pixel, glyph = upper half block.
+        out.push_str(&format!(
+            "\x1b[38;2;{tr};{tg};{tb}m\x1b[48;2;{br};{bg};{bb}m\u{2580}"
+        ));
+    }
+}
+
+/// A magma-style colormap: `t` in `[0,1]` → `(r,g,b)`. Piecewise-linear through
+/// control colours sampled from matplotlib's `magma`.
+fn magma(t: f32) -> (u8, u8, u8) {
+    const STOPS: &[(f32, (f32, f32, f32))] = &[
+        (0.0, (0.0, 0.0, 4.0)),
+        (0.25, (80.0, 18.0, 123.0)),
+        (0.5, (182.0, 54.0, 121.0)),
+        (0.75, (252.0, 136.0, 97.0)),
+        (1.0, (252.0, 253.0, 191.0)),
+    ];
+    let t = t.clamp(0.0, 1.0);
+    for win in STOPS.windows(2) {
+        let (t0, c0) = win[0];
+        let (t1, c1) = win[1];
+        if t <= t1 {
+            let f = if t1 > t0 { (t - t0) / (t1 - t0) } else { 0.0 };
+            return (
+                (c0.0 + (c1.0 - c0.0) * f) as u8,
+                (c0.1 + (c1.1 - c0.1) * f) as u8,
+                (c0.2 + (c1.2 - c0.2) * f) as u8,
+            );
+        }
+    }
+    (252, 253, 191)
+}
+
+/// A horizontal dB colour legend strip.
+fn legend(opts: &RenderOpts) -> String {
+    let mut s = format!("  dB: {:.0}", opts.floor_db);
+    let steps = 32;
+    if opts.ascii {
+        for i in 0..steps {
+            let v = i as f32 / (steps - 1) as f32;
+            let idx = ((v * (ASCII_RAMP.len() - 1) as f32).round() as usize).min(ASCII_RAMP.len() - 1);
+            s.push(ASCII_RAMP[idx] as char);
+        }
+    } else {
+        s.push(' ');
+        for i in 0..steps {
+            let v = i as f32 / (steps - 1) as f32;
+            let (r, g, b) = magma(v);
+            s.push_str(&format!("\x1b[48;2;{r};{g};{b}m "));
+        }
+        s.push_str(RESET);
+    }
+    s.push_str(" 0\n");
+    s
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use crate::dsp::{generators, stft};
+
+    #[test]
+    fn norm_db_maps_floor_and_peak() {
+        // At max magnitude -> 0 dB -> top of range (1.0).
+        assert!((norm_db(1.0, 1.0, -80.0) - 1.0).abs() < 1e-5);
+        // Silence -> 0.0.
+        assert_eq!(norm_db(0.0, 1.0, -80.0), 0.0);
+        // -80 dB (mag 1e-4 of max) -> ~0.0 bottom of range.
+        assert!(norm_db(1e-4, 1.0, -80.0) < 0.01);
+    }
+
+    #[test]
+    fn magma_endpoints() {
+        assert_eq!(magma(0.0), (0, 0, 4));
+        assert_eq!(magma(1.0), (252, 253, 191));
+    }
+
+    #[test]
+    fn render_produces_grid_of_expected_height() {
+        let sig = generators::sine(2000.0, 0.8, 48_000, 48_000);
+        let spec = stft::analyze(&sig, 48_000, 1024, 512);
+        let opts = RenderOpts { width: 40, height: 10, ..Default::default() };
+        let out = render(&spec, &opts);
+        // Header + 10 body rows + time axis (2) + legend = non-trivial.
+        let lines = out.lines().count();
+        assert!(lines >= 13, "expected >=13 lines, got {lines}");
+        assert!(out.contains("Spectrogram"));
+    }
+
+    #[test]
+    fn empty_spectrogram_renders_message() {
+        let spec = stft::analyze(&[], 48_000, 512, 256);
+        let out = render(&spec, &RenderOpts::default());
+        assert!(out.contains("empty"));
+    }
+}
diff --git a/src/dsp/stft.rs b/src/dsp/stft.rs
new file mode 100644
index 0000000..edf778c
--- /dev/null
+++ b/src/dsp/stft.rs
@@ -0,0 +1,105 @@
+//! Short-Time Fourier Transform: slice a signal into overlapping windowed
+//! frames and take each frame's magnitude spectrum. The result is the
+//! time×frequency matrix a spectrogram draws.
+
+use super::fft::real_magnitude_spectrum;
+use super::window::{apply, hann};
+
+/// One STFT analysis: a sequence of magnitude frames plus the geometry needed to
+/// label axes (sample rate, fft size, hop).
+pub struct Spectrogram {
+    /// `frames[t][bin]` = linear magnitude of frequency `bin` at time-step `t`.
+    /// Each inner vec has `fft_size/2 + 1` bins (DC..Nyquist).
+    pub frames: Vec<Vec<f32>>,
+    pub sample_rate: u32,
+    pub fft_size: usize,
+    pub hop: usize,
+}
+
+impl Spectrogram {
+    /// Number of one-sided frequency bins per frame (`fft_size/2 + 1`).
+    pub fn bins(&self) -> usize {
+        self.fft_size / 2 + 1
+    }
+
+    /// Centre frequency (Hz) of bin index `bin`.
+    pub fn bin_hz(&self, bin: usize) -> f32 {
+        bin as f32 * self.sample_rate as f32 / self.fft_size as f32
+    }
+
+    /// Time (seconds) at the start of frame `t`.
+    pub fn frame_time(&self, t: usize) -> f32 {
+        (t * self.hop) as f32 / self.sample_rate as f32
+    }
+}
+
+/// Computes the STFT of `samples` with the given `fft_size` (rounded up to a
+/// power of two) and `hop` (frame advance in samples). A Hann window is applied
+/// to each frame. The final partial frame is zero-padded so trailing audio is
+/// not dropped.
+pub fn analyze(samples: &[f32], sample_rate: u32, fft_size: usize, hop: usize) -> Spectrogram {
+    let fft_size = fft_size.next_power_of_two().max(2);
+    let hop = hop.max(1);
+    let window = hann(fft_size);
+
+    let mut frames = Vec::new();
+    if !samples.is_empty() {
+        let mut start = 0;
+        while start < samples.len() {
+            let end = (start + fft_size).min(samples.len());
+            let mut frame = vec![0.0f32; fft_size];
+            frame[..end - start].copy_from_slice(&samples[start..end]);
+            let windowed = apply(&frame, &window);
+            frames.push(real_magnitude_spectrum(&windowed));
+            start += hop;
+        }
+    }
+
+    Spectrogram {
+        frames,
+        sample_rate,
+        fft_size,
+        hop,
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use std::f64::consts::PI;
+
+    #[test]
+    fn frame_count_follows_hop() {
+        // 1000 samples, hop 250 -> frames start at 0,250,500,750 = 4 frames.
+        let sig = vec![0.0f32; 1000];
+        let s = analyze(&sig, 48_000, 512, 250);
+        assert_eq!(s.frames.len(), 4);
+        assert_eq!(s.bins(), 512 / 2 + 1);
+    }
+
+    #[test]
+    fn tone_lands_in_expected_bin() {
+        // A 3 kHz tone at 48 kHz with a 1024-pt FFT -> bin ≈ 3000/(48000/1024) = 64.
+        let sr = 48_000;
+        let freq = 3000.0;
+        let sig: Vec<f32> = (0..4096)
+            .map(|i| (2.0 * PI * freq * i as f64 / sr as f64).sin() as f32)
+            .collect();
+        let s = analyze(&sig, sr, 1024, 512);
+        let mid = &s.frames[s.frames.len() / 2];
+        let peak_bin = mid
+            .iter()
+            .enumerate()
+            .max_by(|a, b| a.1.partial_cmp(b.1).unwrap())
+            .unwrap()
+            .0;
+        let peak_hz = s.bin_hz(peak_bin);
+        assert!((peak_hz - freq as f32).abs() < 100.0, "peak at {peak_hz} Hz, want {freq}");
+    }
+
+    #[test]
+    fn empty_input_yields_no_frames() {
+        let s = analyze(&[], 48_000, 512, 256);
+        assert!(s.frames.is_empty());
+    }
+}
diff --git a/src/dsp/wav.rs b/src/dsp/wav.rs
new file mode 100644
index 0000000..37d5975
--- /dev/null
+++ b/src/dsp/wav.rs
@@ -0,0 +1,139 @@
+//! A minimal canonical-PCM WAV reader (the analysis-side counterpart to
+//! `audio::recorder::WavWriter`). Handles 16-bit integer PCM — the only format
+//! PeerSpeak writes — in mono or interleaved multi-channel, downmixing to mono
+//! `f32` in `[-1, 1]` for analysis. I/O lives here, at the edge; the rest of the
+//! `dsp` module is pure.
+
+use std::fs;
+use std::path::Path;
+
+/// A decoded WAV: mono `f32` samples plus the original sample rate.
+pub struct WavData {
+    pub samples: Vec<f32>,
+    pub sample_rate: u32,
+}
+
+/// Reads a 16-bit PCM WAV file, downmixing any channels to mono `f32`.
+///
+/// Returns `Err` with a human-readable reason if the file is missing, truncated,
+/// not RIFF/WAVE, or not 16-bit PCM. The parser walks the chunk list rather than
+/// assuming a fixed 44-byte header, so files with extra chunks (`LIST`, `fact`,
+/// …) still read.
+pub fn read(path: &Path) -> Result<WavData, String> {
+    let bytes = fs::read(path).map_err(|e| format!("cannot read {}: {e}", path.display()))?;
+    if bytes.len() < 12 || &bytes[0..4] != b"RIFF" || &bytes[8..12] != b"WAVE" {
+        return Err("not a RIFF/WAVE file".to_string());
+    }
+
+    let mut channels = 0u16;
+    let mut sample_rate = 0u32;
+    let mut bits = 0u16;
+    let mut data: Option<&[u8]> = None;
+
+    // Walk chunks starting after the 12-byte RIFF/WAVE header.
+    let mut pos = 12usize;
+    while pos + 8 <= bytes.len() {
+        let id = &bytes[pos..pos + 4];
+        let size = u32::from_le_bytes([bytes[pos + 4], bytes[pos + 5], bytes[pos + 6], bytes[pos + 7]]) as usize;
+        let body_start = pos + 8;
+        let body_end = (body_start + size).min(bytes.len());
+        match id {
+            b"fmt " if size >= 16 => {
+                let fmt = &bytes[body_start..body_end];
+                let audio_format = u16::from_le_bytes([fmt[0], fmt[1]]);
+                channels = u16::from_le_bytes([fmt[2], fmt[3]]);
+                sample_rate = u32::from_le_bytes([fmt[4], fmt[5], fmt[6], fmt[7]]);
+                bits = u16::from_le_bytes([fmt[14], fmt[15]]);
+                if audio_format != 1 {
+                    return Err(format!("unsupported WAV format tag {audio_format} (need PCM=1)"));
+                }
+            }
+            b"data" => {
+                data = Some(&bytes[body_start..body_end]);
+            }
+            _ => {}
+        }
+        // Chunks are word-aligned: an odd size is padded with one byte.
+        pos = body_start + size + (size & 1);
+    }
+
+    if bits != 16 {
+        return Err(format!("only 16-bit PCM supported, got {bits}-bit"));
+    }
+    let channels = channels.max(1);
+    let data = data.ok_or("no data chunk")?;
+
+    // Interleaved S16LE -> per-frame channel average -> mono f32.
+    let frame_bytes = 2 * channels as usize;
+    let mut samples = Vec::with_capacity(data.len() / frame_bytes.max(1));
+    for frame in data.chunks_exact(frame_bytes) {
+        let mut acc = 0i32;
+        for ch in frame.chunks_exact(2) {
+            acc += i16::from_le_bytes([ch[0], ch[1]]) as i32;
+        }
+        let avg = acc as f32 / channels as f32;
+        samples.push(avg / 32768.0);
+    }
+
+    Ok(WavData { samples, sample_rate })
+}
+
+/// Writes mono `f32` samples (clamped to `[-1, 1]`) as a 16-bit PCM WAV. Used by
+/// the `specview gen` command to materialize conjured test signals.
+pub fn write(path: &Path, samples: &[f32], sample_rate: u32) -> Result<(), String> {
+    let pcm: Vec<i16> = samples
+        .iter()
+        .map(|&s| (s.clamp(-1.0, 1.0) * 32767.0).round() as i16)
+        .collect();
+    let data_bytes = (pcm.len() * 2) as u32;
+    let byte_rate = sample_rate * 2; // mono, 2 bytes/sample
+    let mut out = Vec::with_capacity(44 + pcm.len() * 2);
+    out.extend_from_slice(b"RIFF");
+    out.extend_from_slice(&(36 + data_bytes).to_le_bytes());
+    out.extend_from_slice(b"WAVE");
+    out.extend_from_slice(b"fmt ");
+    out.extend_from_slice(&16u32.to_le_bytes());
+    out.extend_from_slice(&1u16.to_le_bytes()); // PCM
+    out.extend_from_slice(&1u16.to_le_bytes()); // mono
+    out.extend_from_slice(&sample_rate.to_le_bytes());
+    out.extend_from_slice(&byte_rate.to_le_bytes());
+    out.extend_from_slice(&2u16.to_le_bytes()); // block align
+    out.extend_from_slice(&16u16.to_le_bytes()); // bits
+    out.extend_from_slice(b"data");
+    out.extend_from_slice(&data_bytes.to_le_bytes());
+    for s in pcm {
+        out.extend_from_slice(&s.to_le_bytes());
+    }
+    fs::write(path, out).map_err(|e| format!("cannot write {}: {e}", path.display()))
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn write_then_read_round_trips() {
+        let dir = std::env::temp_dir();
+        let path = dir.join(format!("dsp-wav-rt-{}.wav", std::process::id()));
+        let sig: Vec<f32> = (0..1000).map(|i| (i as f32 * 0.01).sin() * 0.5).collect();
+        write(&path, &sig, 48_000).unwrap();
+
+        let back = read(&path).unwrap();
+        assert_eq!(back.sample_rate, 48_000);
+        assert_eq!(back.samples.len(), sig.len());
+        // 16-bit quantization tolerance.
+        for (a, b) in sig.iter().zip(&back.samples) {
+            assert!((a - b).abs() < 1e-3, "{a} vs {b}");
+        }
+        let _ = fs::remove_file(&path);
+    }
+
+    #[test]
+    fn rejects_non_wave() {
+        let dir = std::env::temp_dir();
+        let path = dir.join(format!("dsp-wav-bad-{}.bin", std::process::id()));
+        fs::write(&path, b"not a wav at all").unwrap();
+        assert!(read(&path).is_err());
+        let _ = fs::remove_file(&path);
+    }
+}
diff --git a/src/dsp/window.rs b/src/dsp/window.rs
new file mode 100644
index 0000000..85a3685
--- /dev/null
+++ b/src/dsp/window.rs
@@ -0,0 +1,55 @@
+//! Analysis windows for the STFT.
+//!
+//! A raw rectangular frame leaks spectral energy across bins (the abrupt edges
+//! look like discontinuities to the FFT). A Hann window tapers each frame to
+//! zero at its edges, trading a little main-lobe width for much lower side-lobe
+//! leakage — the standard choice for a spectrogram.
+
+use std::f64::consts::PI;
+
+/// Returns a length-`n` periodic Hann window, `w[i] = 0.5·(1 - cos(2πi/n))`.
+///
+/// The *periodic* form (denominator `n`, not `n-1`) is used because STFT frames
+/// tile the signal; it gives perfect overlap-add reconstruction at 50% hop.
+pub fn hann(n: usize) -> Vec<f32> {
+    if n <= 1 {
+        return vec![1.0; n];
+    }
+    (0..n)
+        .map(|i| (0.5 - 0.5 * (2.0 * PI * i as f64 / n as f64).cos()) as f32)
+        .collect()
+}
+
+/// Applies `window` to `frame` element-wise into a new buffer. Lengths must match.
+pub fn apply(frame: &[f32], window: &[f32]) -> Vec<f32> {
+    debug_assert_eq!(frame.len(), window.len());
+    frame.iter().zip(window).map(|(&s, &w)| s * w).collect()
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn hann_endpoints_are_zero_and_center_is_one() {
+        let w = hann(8);
+        assert!(w[0].abs() < 1e-6, "first sample tapers to ~0, got {}", w[0]);
+        // Periodic Hann peaks at the midpoint n/2.
+        assert!((w[4] - 1.0).abs() < 1e-6, "center peaks at 1, got {}", w[4]);
+    }
+
+    #[test]
+    fn hann_is_symmetric_about_center() {
+        let w = hann(16);
+        // Periodic window is symmetric across indices 1..n-1.
+        for i in 1..8 {
+            assert!((w[i] - w[16 - i]).abs() < 1e-6);
+        }
+    }
+
+    #[test]
+    fn apply_scales_samples() {
+        let out = apply(&[2.0, 2.0], &[0.5, 0.25]);
+        assert_eq!(out, vec![1.0, 0.5]);
+    }
+}
diff --git a/src/lib.rs b/src/lib.rs
index c85a211..77c8fa3 100644
--- a/src/lib.rs
+++ b/src/lib.rs
@@ -1,5 +1,6 @@
 pub mod audio;
 pub mod codec;
+pub mod dsp;
 pub mod network;
 pub mod core;
 pub mod app;