feat: Android VoIP client — Phase 1 (audio quality, network adaptation, crate skeleton)

- New wzp-android crate with Oboe C++ backend, lock-free SPSC ring buffers,
  engine orchestrator, codec pipeline, and Android Gradle project structure
- AEC (NLMS adaptive filter), AGC (two-stage with fast attack/slow release),
  windowed-sinc FIR resampler replacing linear interpolation (wzp-codec)
- Opus encoder tuning: complexity 7 default, set_expected_loss support
- Mobile jitter buffer: asymmetric EMA (fast up/slow down), handoff spike
  detection with 2s cooldown, configurable safety margin
- Network-aware quality control: cellular-specific thresholds, faster
  downgrade on cellular, proactive tier drop on WiFi→cellular handoff,
  FEC ratio boost during network transitions
- Handoff detection in PathMonitor via RTT jitter spike analysis

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

crates/wzp-codec/src/adaptive.rs

@@ -14,7 +14,7 @@ use crate::codec2_dec::Codec2Decoder;
 use crate::codec2_enc::Codec2Encoder;
 use crate::opus_dec::OpusDecoder;
 use crate::opus_enc::OpusEncoder;
-use crate::resample;
+use crate::resample::{Downsampler48to8, Upsampler8to48};
 
 // ─── Helpers ─────────────────────────────────────────────────────────────────
 
@@ -54,6 +54,7 @@ pub struct AdaptiveEncoder {
     opus: OpusEncoder,
     codec2: Codec2Encoder,
     active: CodecId,
+    downsampler: Downsampler48to8,
 }
 
 impl AdaptiveEncoder {
@@ -66,6 +67,7 @@ impl AdaptiveEncoder {
             opus,
             codec2,
             active: profile.codec,
+            downsampler: Downsampler48to8::new(),
         })
     }
 }
@@ -74,7 +76,7 @@ impl AudioEncoder for AdaptiveEncoder {
     fn encode(&mut self, pcm: &[i16], out: &mut [u8]) -> Result<usize, CodecError> {
         if is_codec2(self.active) {
             // Downsample 48 kHz → 8 kHz then encode via Codec2.
-            let pcm_8k = resample::resample_48k_to_8k(pcm);
+            let pcm_8k = self.downsampler.process(pcm);
             self.codec2.encode(&pcm_8k, out)
         } else {
             self.opus.encode(pcm, out)
@@ -126,6 +128,7 @@ pub struct AdaptiveDecoder {
     opus: OpusDecoder,
     codec2: Codec2Decoder,
     active: CodecId,
+    upsampler: Upsampler8to48,
 }
 
 impl AdaptiveDecoder {
@@ -138,6 +141,7 @@ impl AdaptiveDecoder {
             opus,
             codec2,
             active: profile.codec,
+            upsampler: Upsampler8to48::new(),
         })
     }
 }
@@ -149,7 +153,7 @@ impl AudioDecoder for AdaptiveDecoder {
             let c2_samples = self.codec2_frame_samples();
             let mut buf_8k = vec![0i16; c2_samples];
             let n = self.codec2.decode(encoded, &mut buf_8k)?;
-            let pcm_48k = resample::resample_8k_to_48k(&buf_8k[..n]);
+            let pcm_48k = self.upsampler.process(&buf_8k[..n]);
             let out_len = pcm_48k.len().min(pcm.len());
             pcm[..out_len].copy_from_slice(&pcm_48k[..out_len]);
             Ok(out_len)
@@ -163,7 +167,7 @@ impl AudioDecoder for AdaptiveDecoder {
             let c2_samples = self.codec2_frame_samples();
             let mut buf_8k = vec![0i16; c2_samples];
             let n = self.codec2.decode_lost(&mut buf_8k)?;
-            let pcm_48k = resample::resample_8k_to_48k(&buf_8k[..n]);
+            let pcm_48k = self.upsampler.process(&buf_8k[..n]);
             let out_len = pcm_48k.len().min(pcm.len());
             pcm[..out_len].copy_from_slice(&pcm_48k[..out_len]);
             Ok(out_len)

crates/wzp-codec/src/aec.rs

@@ -0,0 +1,228 @@
+//! Acoustic Echo Cancellation using NLMS adaptive filter.
+//! Processes 480-sample (10ms) sub-frames at 48kHz.
+
+/// NLMS (Normalized Least Mean Squares) adaptive filter echo canceller.
+///
+/// Removes acoustic echo by modelling the echo path between the far-end
+/// (speaker) signal and the near-end (microphone) signal, then subtracting
+/// the estimated echo from the near-end in real time.
+pub struct EchoCanceller {
+    filter_coeffs: Vec<f32>,
+    filter_len: usize,
+    far_end_buf: Vec<f32>,
+    far_end_pos: usize,
+    mu: f32,
+    enabled: bool,
+}
+
+impl EchoCanceller {
+    /// Create a new echo canceller.
+    ///
+    /// * `sample_rate` — typically 48000
+    /// * `filter_ms`   — echo-tail length in milliseconds (e.g. 100 for 100 ms)
+    pub fn new(sample_rate: u32, filter_ms: u32) -> Self {
+        let filter_len = (sample_rate as usize) * (filter_ms as usize) / 1000;
+        Self {
+            filter_coeffs: vec![0.0f32; filter_len],
+            filter_len,
+            far_end_buf: vec![0.0f32; filter_len],
+            far_end_pos: 0,
+            mu: 0.01,
+            enabled: true,
+        }
+    }
+
+    /// Feed far-end (speaker/playback) samples into the circular buffer.
+    ///
+    /// Must be called with the audio that was played out through the speaker
+    /// *before* the corresponding near-end frame is processed.
+    pub fn feed_farend(&mut self, farend: &[i16]) {
+        for &s in farend {
+            self.far_end_buf[self.far_end_pos] = s as f32;
+            self.far_end_pos = (self.far_end_pos + 1) % self.filter_len;
+        }
+    }
+
+    /// Process a near-end (microphone) frame, removing the estimated echo.
+    ///
+    /// Returns the echo-return-loss enhancement (ERLE) as a ratio: the RMS of
+    /// the original near-end divided by the RMS of the residual.  Values > 1.0
+    /// mean echo was reduced.
+    pub fn process_frame(&mut self, nearend: &mut [i16]) -> f32 {
+        if !self.enabled {
+            return 1.0;
+        }
+
+        let n = nearend.len();
+        let fl = self.filter_len;
+
+        let mut sum_near_sq: f64 = 0.0;
+        let mut sum_err_sq: f64 = 0.0;
+
+        for i in 0..n {
+            let near_f = nearend[i] as f32;
+
+            // --- estimate echo as dot(coeffs, farend_window) ---
+            // The far-end window for this sample starts at
+            //   (far_end_pos - 1 - i) mod filter_len   (most recent)
+            // and goes back filter_len samples.
+            let mut echo_est: f32 = 0.0;
+            let mut power: f32 = 0.0;
+
+            // Position of the most-recent far-end sample for this near-end sample.
+            // far_end_pos points to the *next write* position, so the most-recent
+            // sample written is at far_end_pos - 1.  We have already called
+            // feed_farend for this block, so the relevant samples are the last
+            // filter_len entries ending just before the current write position,
+            // offset by how far we are into this near-end frame.
+            //
+            // For sample i of the near-end frame, the corresponding far-end
+            // "now" is far_end_pos - n + i  (wrapping).
+            // far_end_pos points to next-write, so most recent sample is at
+            // far_end_pos - 1.  For the i-th near-end sample we want the
+            // far-end "now" to be at (far_end_pos - n + i).  We add fl
+            // repeatedly to avoid underflow on the usize subtraction.
+            let base = (self.far_end_pos + fl * ((n / fl) + 2) + i - n) % fl;
+
+            for k in 0..fl {
+                let fe_idx = (base + fl - k) % fl;
+                let fe = self.far_end_buf[fe_idx];
+                echo_est += self.filter_coeffs[k] * fe;
+                power += fe * fe;
+            }
+
+            let error = near_f - echo_est;
+
+            // --- NLMS coefficient update ---
+            let norm = power + 1.0; // +1 regularisation to avoid div-by-zero
+            let step = self.mu * error / norm;
+
+            for k in 0..fl {
+                let fe_idx = (base + fl - k) % fl;
+                let fe = self.far_end_buf[fe_idx];
+                self.filter_coeffs[k] += step * fe;
+            }
+
+            // Clamp output
+            let out = error.max(-32768.0).min(32767.0);
+            nearend[i] = out as i16;
+
+            sum_near_sq += (near_f as f64) * (near_f as f64);
+            sum_err_sq += (out as f64) * (out as f64);
+        }
+
+        // ERLE ratio
+        if sum_err_sq < 1.0 {
+            return 100.0; // near-perfect cancellation
+        }
+        (sum_near_sq / sum_err_sq).sqrt() as f32
+    }
+
+    /// Enable or disable echo cancellation.
+    pub fn set_enabled(&mut self, enabled: bool) {
+        self.enabled = enabled;
+    }
+
+    /// Returns whether echo cancellation is currently enabled.
+    pub fn is_enabled(&self) -> bool {
+        self.enabled
+    }
+
+    /// Reset the adaptive filter to its initial state.
+    ///
+    /// Zeroes out all filter coefficients and the far-end circular buffer.
+    pub fn reset(&mut self) {
+        self.filter_coeffs.iter_mut().for_each(|c| *c = 0.0);
+        self.far_end_buf.iter_mut().for_each(|s| *s = 0.0);
+        self.far_end_pos = 0;
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn aec_creates_with_correct_filter_len() {
+        let aec = EchoCanceller::new(48000, 100);
+        assert_eq!(aec.filter_len, 4800);
+        assert_eq!(aec.filter_coeffs.len(), 4800);
+        assert_eq!(aec.far_end_buf.len(), 4800);
+    }
+
+    #[test]
+    fn aec_passthrough_when_disabled() {
+        let mut aec = EchoCanceller::new(48000, 100);
+        aec.set_enabled(false);
+        assert!(!aec.is_enabled());
+
+        let original: Vec<i16> = (0..480).map(|i| (i * 10) as i16).collect();
+        let mut frame = original.clone();
+        let erle = aec.process_frame(&mut frame);
+        assert_eq!(erle, 1.0);
+        assert_eq!(frame, original);
+    }
+
+    #[test]
+    fn aec_reset_zeroes_state() {
+        let mut aec = EchoCanceller::new(48000, 10); // short for test speed
+        let farend: Vec<i16> = (0..480).map(|i| ((i * 37) % 1000) as i16).collect();
+        aec.feed_farend(&farend);
+
+        aec.reset();
+
+        assert!(aec.filter_coeffs.iter().all(|&c| c == 0.0));
+        assert!(aec.far_end_buf.iter().all(|&s| s == 0.0));
+        assert_eq!(aec.far_end_pos, 0);
+    }
+
+    #[test]
+    fn aec_reduces_echo_of_known_signal() {
+        // Use a small filter for speed.  Feed a known far-end signal, then
+        // present the *same* signal as near-end (perfect echo, no room).
+        // After adaptation the output energy should drop.
+        let filter_ms = 5; // 240 taps at 48 kHz
+        let mut aec = EchoCanceller::new(48000, filter_ms);
+
+        // Generate a simple repeating pattern.
+        let frame_len = 480usize;
+        let make_frame = |offset: usize| -> Vec<i16> {
+            (0..frame_len)
+                .map(|i| {
+                    let t = (offset + i) as f64 / 48000.0;
+                    (5000.0 * (2.0 * std::f64::consts::PI * 300.0 * t).sin()) as i16
+                })
+                .collect()
+        };
+
+        // Warm up the adaptive filter with several frames.
+        let mut last_erle = 1.0f32;
+        for frame_idx in 0..40 {
+            let farend = make_frame(frame_idx * frame_len);
+            aec.feed_farend(&farend);
+
+            // Near-end = exact copy of far-end (pure echo).
+            let mut nearend = farend.clone();
+            last_erle = aec.process_frame(&mut nearend);
+        }
+
+        // After 40 frames the ERLE should be meaningfully > 1.
+        assert!(
+            last_erle > 1.0,
+            "expected ERLE > 1.0 after adaptation, got {last_erle}"
+        );
+    }
+
+    #[test]
+    fn aec_silence_passthrough() {
+        let mut aec = EchoCanceller::new(48000, 10);
+        // Feed silence far-end
+        aec.feed_farend(&vec![0i16; 480]);
+        // Near-end is silence too
+        let mut frame = vec![0i16; 480];
+        let erle = aec.process_frame(&mut frame);
+        assert!(erle >= 1.0);
+        // Output should still be silence
+        assert!(frame.iter().all(|&s| s == 0));
+    }
+}

crates/wzp-codec/src/agc.rs

@@ -0,0 +1,219 @@
+//! Automatic Gain Control (AGC) with two-stage smoothing.
+//!
+//! Uses a fast attack / slow release envelope follower to keep the
+//! output signal near a configurable target RMS level.  This prevents
+//! both clipping (when the speaker is too loud) and inaudibility (when
+//! the speaker is too quiet or far from the mic).
+
+/// Two-stage automatic gain control.
+///
+/// The gain is adjusted per-frame based on the measured RMS energy,
+/// with a fast attack (gain decreases quickly when signal gets louder)
+/// and a slow release (gain increases gradually when signal gets quieter).
+pub struct AutoGainControl {
+    target_rms: f64,
+    current_gain: f64,
+    min_gain: f64,
+    max_gain: f64,
+    attack_alpha: f64,
+    release_alpha: f64,
+    enabled: bool,
+}
+
+impl AutoGainControl {
+    /// Create a new AGC with sensible VoIP defaults.
+    pub fn new() -> Self {
+        Self {
+            target_rms: 3000.0,   // ~-20 dBFS for i16
+            current_gain: 1.0,
+            min_gain: 0.5,
+            max_gain: 32.0,
+            attack_alpha: 0.3,    // fast attack
+            release_alpha: 0.02,  // slow release
+            enabled: true,
+        }
+    }
+
+    /// Process a frame of PCM audio in-place, applying gain adjustment.
+    pub fn process_frame(&mut self, pcm: &mut [i16]) {
+        if !self.enabled {
+            return;
+        }
+
+        // Compute RMS of the frame.
+        let rms = Self::compute_rms(pcm);
+
+        // Don't amplify near-silence — it would just boost noise.
+        if rms < 10.0 {
+            return;
+        }
+
+        // Desired instantaneous gain.
+        let desired_gain = (self.target_rms / rms).clamp(self.min_gain, self.max_gain);
+
+        // Smooth the gain transition.
+        let alpha = if desired_gain < self.current_gain {
+            // Signal is louder than target → reduce gain quickly (attack).
+            self.attack_alpha
+        } else {
+            // Signal is quieter than target → raise gain slowly (release).
+            self.release_alpha
+        };
+
+        self.current_gain = self.current_gain * (1.0 - alpha) + desired_gain * alpha;
+
+        // Apply gain to each sample with hard limiting at ±31000 (~0.946 * i16::MAX).
+        const LIMIT: f64 = 31000.0;
+        let gain = self.current_gain;
+        for sample in pcm.iter_mut() {
+            let amplified = (*sample as f64) * gain;
+            let clamped = amplified.clamp(-LIMIT, LIMIT);
+            *sample = clamped as i16;
+        }
+    }
+
+    /// Enable or disable the AGC.
+    pub fn set_enabled(&mut self, enabled: bool) {
+        self.enabled = enabled;
+    }
+
+    /// Returns whether the AGC is currently enabled.
+    pub fn is_enabled(&self) -> bool {
+        self.enabled
+    }
+
+    /// Current gain expressed in dB.
+    pub fn current_gain_db(&self) -> f64 {
+        20.0 * self.current_gain.log10()
+    }
+
+    /// Compute the RMS (root mean square) of a PCM buffer.
+    fn compute_rms(pcm: &[i16]) -> f64 {
+        if pcm.is_empty() {
+            return 0.0;
+        }
+        let sum_sq: f64 = pcm.iter().map(|&s| (s as f64) * (s as f64)).sum();
+        (sum_sq / pcm.len() as f64).sqrt()
+    }
+}
+
+impl Default for AutoGainControl {
+    fn default() -> Self {
+        Self::new()
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn agc_creates_with_defaults() {
+        let agc = AutoGainControl::new();
+        assert!(agc.is_enabled());
+        assert!((agc.current_gain - 1.0).abs() < f64::EPSILON);
+    }
+
+    #[test]
+    fn agc_passthrough_when_disabled() {
+        let mut agc = AutoGainControl::new();
+        agc.set_enabled(false);
+
+        let original: Vec<i16> = (0..960).map(|i| (i * 5) as i16).collect();
+        let mut frame = original.clone();
+        agc.process_frame(&mut frame);
+
+        assert_eq!(frame, original);
+    }
+
+    #[test]
+    fn agc_does_not_amplify_silence() {
+        let mut agc = AutoGainControl::new();
+        let mut frame = vec![0i16; 960];
+        agc.process_frame(&mut frame);
+        assert!(frame.iter().all(|&s| s == 0));
+        // Gain should remain at initial value.
+        assert!((agc.current_gain - 1.0).abs() < f64::EPSILON);
+    }
+
+    #[test]
+    fn agc_amplifies_quiet_signal() {
+        let mut agc = AutoGainControl::new();
+
+        // Very quiet signal (RMS ~ 50).
+        let mut frame: Vec<i16> = (0..960)
+            .map(|i| {
+                let t = i as f64 / 48000.0;
+                (50.0 * (2.0 * std::f64::consts::PI * 440.0 * t).sin()) as i16
+            })
+            .collect();
+
+        // Process several frames to let the gain ramp up.
+        for _ in 0..50 {
+            let mut f = frame.clone();
+            agc.process_frame(&mut f);
+            frame = f;
+        }
+
+        // Gain should have increased past 1.0.
+        assert!(
+            agc.current_gain > 1.05,
+            "expected gain > 1.05 for quiet signal, got {}",
+            agc.current_gain
+        );
+    }
+
+    #[test]
+    fn agc_attenuates_loud_signal() {
+        let mut agc = AutoGainControl::new();
+
+        // Loud signal (RMS ~ 20000).
+        let frame: Vec<i16> = (0..960)
+            .map(|i| {
+                let t = i as f64 / 48000.0;
+                (28000.0 * (2.0 * std::f64::consts::PI * 440.0 * t).sin()) as i16
+            })
+            .collect();
+
+        // Process several frames.
+        for _ in 0..20 {
+            let mut f = frame.clone();
+            agc.process_frame(&mut f);
+        }
+
+        // Gain should have decreased below 1.0.
+        assert!(
+            agc.current_gain < 1.0,
+            "expected gain < 1.0 for loud signal, got {}",
+            agc.current_gain
+        );
+    }
+
+    #[test]
+    fn agc_output_within_limits() {
+        let mut agc = AutoGainControl::new();
+        // Force a high gain by processing many quiet frames first.
+        for _ in 0..100 {
+            let mut f: Vec<i16> = vec![100; 960];
+            agc.process_frame(&mut f);
+        }
+
+        // Now send a louder frame — output should still be within ±31000.
+        let mut frame: Vec<i16> = vec![20000; 960];
+        agc.process_frame(&mut frame);
+        assert!(
+            frame.iter().all(|&s| s.abs() <= 31000),
+            "output samples must be within ±31000"
+        );
+    }
+
+    #[test]
+    fn agc_gain_db_at_unity() {
+        let agc = AutoGainControl::new();
+        let db = agc.current_gain_db();
+        assert!(
+            db.abs() < 0.01,
+            "expected ~0 dB at unity gain, got {db}"
+        );
+    }
+}

crates/wzp-codec/src/lib.rs

@@ -10,6 +10,8 @@
 //! trait-object encoders/decoders that handle adaptive switching internally.
 
 pub mod adaptive;
+pub mod aec;
+pub mod agc;
 pub mod codec2_dec;
 pub mod codec2_enc;
 pub mod denoise;
@@ -19,6 +21,8 @@ pub mod resample;
 pub mod silence;
 
 pub use adaptive::{AdaptiveDecoder, AdaptiveEncoder};
+pub use aec::EchoCanceller;
+pub use agc::AutoGainControl;
 pub use denoise::NoiseSupressor;
 pub use silence::{ComfortNoise, SilenceDetector};
 pub use wzp_proto::{AudioDecoder, AudioEncoder, CodecId, QualityProfile};

crates/wzp-codec/src/opus_enc.rs

@@ -40,6 +40,11 @@ impl OpusEncoder {
             .set_signal(Signal::Voice)
             .map_err(|e| CodecError::EncodeFailed(format!("set signal: {e}")))?;
 
+        // Default complexity 7 — good quality/CPU trade-off for VoIP
+        enc.inner
+            .set_complexity(7)
+            .map_err(|e| CodecError::EncodeFailed(format!("set complexity: {e}")))?;
+
         Ok(enc)
     }
 
@@ -56,6 +61,21 @@ impl OpusEncoder {
     pub fn frame_samples(&self) -> usize {
         (48_000 * self.frame_duration_ms as usize) / 1000
     }
+
+    /// Set the encoder complexity (0-10). Higher values produce better quality
+    /// at the cost of more CPU. Default is 7.
+    pub fn set_complexity(&mut self, complexity: i32) {
+        let c = (complexity as u8).min(10);
+        let _ = self.inner.set_complexity(c);
+    }
+
+    /// Hint the encoder about expected packet loss percentage (0-100).
+    ///
+    /// Higher values cause the encoder to use more redundancy to survive
+    /// packet loss, at the expense of slightly higher bitrate.
+    pub fn set_expected_loss(&mut self, loss_pct: u8) {
+        let _ = self.inner.set_packet_loss_perc(loss_pct.min(100));
+    }
 }
 
 impl AudioEncoder for OpusEncoder {

crates/wzp-codec/src/resample.rs

@@ -1,55 +1,258 @@
-//! Simple linear resampler for 48 kHz <-> 8 kHz conversion.
+//! Windowed-sinc FIR resampler for 48 kHz <-> 8 kHz conversion.
 //!
-//! These are basic implementations suitable for voice. For higher quality,
-//! replace with the `rubato` crate later.
+//! Provides both stateless free functions (backward-compatible) and stateful
+//! `Downsampler48to8` / `Upsampler8to48` structs that maintain overlap history
+//! between frames for glitch-free streaming.
 
-/// Downsample from 48 kHz to 8 kHz (6:1 decimation with averaging).
+use std::f64::consts::PI;
+
+// ─── FIR kernel parameters ─────────────────────────────────────────────────
+
+/// Number of FIR taps in the anti-alias / interpolation filter.
+const FIR_TAPS: usize = 48;
+/// Kaiser window beta parameter — controls sidelobe attenuation.
+const KAISER_BETA: f64 = 8.0;
+/// Cutoff frequency in Hz for the low-pass filter (just below 4 kHz Nyquist of 8 kHz).
+const CUTOFF_HZ: f64 = 3800.0;
+/// Working sample rate in Hz.
+const SAMPLE_RATE: f64 = 48000.0;
+/// Decimation / interpolation ratio between 48 kHz and 8 kHz.
+const RATIO: usize = 6;
+
+// ─── Kaiser window helpers ─────────────────────────────────────────────────
+
+/// Zeroth-order modified Bessel function of the first kind, I₀(x).
 ///
-/// Each output sample is the average of 6 consecutive input samples,
-/// providing basic anti-aliasing via a box filter.
-pub fn resample_48k_to_8k(input: &[i16]) -> Vec<i16> {
-    const RATIO: usize = 6;
-    let out_len = input.len() / RATIO;
-    let mut output = Vec::with_capacity(out_len);
-
-    for chunk in input.chunks_exact(RATIO) {
-        let sum: i32 = chunk.iter().map(|&s| s as i32).sum();
-        output.push((sum / RATIO as i32) as i16);
+/// Computed via the well-known power-series expansion, converging rapidly
+/// for the moderate values of x used in Kaiser window design.
+fn bessel_i0(x: f64) -> f64 {
+    let mut sum = 1.0f64;
+    let mut term = 1.0f64;
+    let half_x = x / 2.0;
+    for k in 1..=25 {
+        term *= (half_x / k as f64) * (half_x / k as f64);
+        sum += term;
+        if term < 1e-12 * sum {
+            break;
+        }
     }
-
-    output
+    sum
 }
 
-/// Upsample from 8 kHz to 48 kHz (1:6 interpolation with linear interp).
+/// Build a windowed-sinc low-pass FIR kernel.
 ///
-/// Linearly interpolates between each pair of input samples to produce
-/// 6 output samples per input sample.
-pub fn resample_8k_to_48k(input: &[i16]) -> Vec<i16> {
-    const RATIO: usize = 6;
-    if input.is_empty() {
-        return Vec::new();
-    }
+/// Returns `FIR_TAPS` coefficients normalised so that the DC gain is exactly 1.0.
+fn build_fir_kernel() -> [f64; FIR_TAPS] {
+    let mut kernel = [0.0f64; FIR_TAPS];
+    let m = (FIR_TAPS - 1) as f64;
+    let fc = CUTOFF_HZ / SAMPLE_RATE; // normalised cutoff (0..0.5)
+    let beta_denom = bessel_i0(KAISER_BETA);
 
-    let out_len = input.len() * RATIO;
-    let mut output = Vec::with_capacity(out_len);
-
-    for i in 0..input.len() {
-        let current = input[i] as i32;
-        let next = if i + 1 < input.len() {
-            input[i + 1] as i32
+    for i in 0..FIR_TAPS {
+        // Sinc
+        let n = i as f64 - m / 2.0;
+        let sinc = if n.abs() < 1e-12 {
+            2.0 * fc
         } else {
-            current // hold last sample
+            (2.0 * PI * fc * n).sin() / (PI * n)
         };
 
-        for j in 0..RATIO {
-            let interp = current + (next - current) * j as i32 / RATIO as i32;
-            output.push(interp as i16);
+        // Kaiser window
+        let t = 2.0 * i as f64 / m - 1.0; // range [-1, 1]
+        let kaiser = bessel_i0(KAISER_BETA * (1.0 - t * t).max(0.0).sqrt()) / beta_denom;
+
+        kernel[i] = sinc * kaiser;
+    }
+
+    // Normalise to unity DC gain.
+    let sum: f64 = kernel.iter().sum();
+    if sum.abs() > 1e-15 {
+        for k in kernel.iter_mut() {
+            *k /= sum;
         }
     }
 
-    output
+    kernel
 }
 
+// ─── Stateful Downsampler 48→8 ─────────────────────────────────────────────
+
+/// Stateful polyphase FIR downsampler from 48 kHz to 8 kHz.
+///
+/// Maintains `FIR_TAPS - 1` samples of history between successive calls to
+/// `process()` for seamless frame boundaries.
+pub struct Downsampler48to8 {
+    kernel: [f64; FIR_TAPS],
+    history: Vec<f64>,
+}
+
+impl Downsampler48to8 {
+    pub fn new() -> Self {
+        Self {
+            kernel: build_fir_kernel(),
+            history: vec![0.0; FIR_TAPS - 1],
+        }
+    }
+
+    /// Downsample a block of 48 kHz samples to 8 kHz.
+    ///
+    /// The input length should be a multiple of 6; any trailing samples that
+    /// don't form a complete output sample are consumed into the history.
+    pub fn process(&mut self, input: &[i16]) -> Vec<i16> {
+        let hist_len = self.history.len(); // FIR_TAPS - 1
+        let total_len = hist_len + input.len();
+
+        // Build a working buffer: history ++ input (as f64).
+        let mut work = Vec::with_capacity(total_len);
+        work.extend_from_slice(&self.history);
+        work.extend(input.iter().map(|&s| s as f64));
+
+        let out_len = input.len() / RATIO;
+        let mut output = Vec::with_capacity(out_len);
+
+        for i in 0..out_len {
+            // The centre of the filter for output sample i sits at
+            // position hist_len + i*RATIO in the work buffer (aligning
+            // with the first new input sample at decimation phase 0).
+            let centre = hist_len + i * RATIO;
+            let start = centre + 1 - FIR_TAPS; // may be 0 for the first few
+
+            let mut acc = 0.0f64;
+            for k in 0..FIR_TAPS {
+                let idx = start + k;
+                if idx < work.len() {
+                    acc += work[idx] * self.kernel[k];
+                }
+            }
+            output.push(acc.round().clamp(-32768.0, 32767.0) as i16);
+        }
+
+        // Update history: keep the last (FIR_TAPS - 1) samples from work.
+        if work.len() >= hist_len {
+            self.history
+                .copy_from_slice(&work[work.len() - hist_len..]);
+        } else {
+            // Input was shorter than history — shift.
+            let shift = hist_len - work.len();
+            self.history.copy_within(shift.., 0);
+            for (i, &v) in work.iter().enumerate() {
+                self.history[hist_len - work.len() + i] = v;
+            }
+        }
+
+        output
+    }
+}
+
+impl Default for Downsampler48to8 {
+    fn default() -> Self {
+        Self::new()
+    }
+}
+
+// ─── Stateful Upsampler 8→48 ───────────────────────────────────────────────
+
+/// Stateful FIR upsampler from 8 kHz to 48 kHz.
+///
+/// Inserts zeros between input samples (zero-stuffing), then applies the
+/// low-pass FIR to remove imaging, with gain compensation of `RATIO`.
+pub struct Upsampler8to48 {
+    kernel: [f64; FIR_TAPS],
+    history: Vec<f64>,
+}
+
+impl Upsampler8to48 {
+    pub fn new() -> Self {
+        Self {
+            kernel: build_fir_kernel(),
+            history: vec![0.0; FIR_TAPS - 1],
+        }
+    }
+
+    /// Upsample a block of 8 kHz samples to 48 kHz.
+    pub fn process(&mut self, input: &[i16]) -> Vec<i16> {
+        let hist_len = self.history.len(); // FIR_TAPS - 1
+
+        // Zero-stuff: insert RATIO-1 zeros between each input sample.
+        let stuffed_len = input.len() * RATIO;
+        let total_len = hist_len + stuffed_len;
+
+        let mut work = Vec::with_capacity(total_len);
+        work.extend_from_slice(&self.history);
+        for &s in input {
+            work.push(s as f64);
+            for _ in 1..RATIO {
+                work.push(0.0);
+            }
+        }
+
+        let out_len = stuffed_len;
+        let mut output = Vec::with_capacity(out_len);
+
+        // The gain factor compensates for the zeros introduced by stuffing.
+        let gain = RATIO as f64;
+
+        for i in 0..out_len {
+            let centre = hist_len + i;
+            let start = centre + 1 - FIR_TAPS;
+
+            let mut acc = 0.0f64;
+            for k in 0..FIR_TAPS {
+                let idx = start + k;
+                if idx < work.len() {
+                    acc += work[idx] * self.kernel[k];
+                }
+            }
+            acc *= gain;
+            output.push(acc.round().clamp(-32768.0, 32767.0) as i16);
+        }
+
+        // Update history.
+        if work.len() >= hist_len {
+            self.history
+                .copy_from_slice(&work[work.len() - hist_len..]);
+        } else {
+            let shift = hist_len - work.len();
+            self.history.copy_within(shift.., 0);
+            for (i, &v) in work.iter().enumerate() {
+                self.history[hist_len - work.len() + i] = v;
+            }
+        }
+
+        output
+    }
+}
+
+impl Default for Upsampler8to48 {
+    fn default() -> Self {
+        Self::new()
+    }
+}
+
+// ─── Backward-compatible free functions ─────────────────────────────────────
+
+/// Downsample from 48 kHz to 8 kHz (6:1 decimation with FIR anti-alias filter).
+///
+/// This is a convenience wrapper that creates a temporary [`Downsampler48to8`].
+/// For streaming use, prefer the stateful struct to avoid edge artefacts between
+/// frames.
+pub fn resample_48k_to_8k(input: &[i16]) -> Vec<i16> {
+    let mut ds = Downsampler48to8::new();
+    ds.process(input)
+}
+
+/// Upsample from 8 kHz to 48 kHz (1:6 interpolation with FIR imaging filter).
+///
+/// This is a convenience wrapper that creates a temporary [`Upsampler8to48`].
+/// For streaming use, prefer the stateful struct to avoid edge artefacts between
+/// frames.
+pub fn resample_8k_to_48k(input: &[i16]) -> Vec<i16> {
+    let mut us = Upsampler8to48::new();
+    us.process(input)
+}
+
+// ─── Tests ──────────────────────────────────────────────────────────────────
+
 #[cfg(test)]
 mod tests {
     use super::*;
@@ -66,12 +269,28 @@ mod tests {
 
     #[test]
     fn dc_signal_preserved() {
-        // A constant signal should survive resampling
+        // A constant signal should survive resampling (approximately).
         let input = vec![1000i16; 960];
         let down = resample_48k_to_8k(&input);
-        assert!(down.iter().all(|&s| s == 1000));
+        // Allow some edge transient — check that the middle samples are close.
+        let mid_start = down.len() / 4;
+        let mid_end = 3 * down.len() / 4;
+        for &s in &down[mid_start..mid_end] {
+            assert!(
+                (s - 1000).abs() < 50,
+                "DC downsampled sample {s} too far from 1000"
+            );
+        }
+
         let up = resample_8k_to_48k(&down);
-        assert!(up.iter().all(|&s| s == 1000));
+        let mid_start_up = up.len() / 4;
+        let mid_end_up = 3 * up.len() / 4;
+        for &s in &up[mid_start_up..mid_end_up] {
+            assert!(
+                (s - 1000).abs() < 100,
+                "DC upsampled sample {s} too far from 1000"
+            );
+        }
     }
 
     #[test]
@@ -79,4 +298,40 @@ mod tests {
         assert!(resample_48k_to_8k(&[]).is_empty());
         assert!(resample_8k_to_48k(&[]).is_empty());
     }
+
+    #[test]
+    fn stateful_downsampler_produces_correct_length() {
+        let mut ds = Downsampler48to8::new();
+        let out = ds.process(&vec![0i16; 960]);
+        assert_eq!(out.len(), 160);
+        let out2 = ds.process(&vec![0i16; 960]);
+        assert_eq!(out2.len(), 160);
+    }
+
+    #[test]
+    fn stateful_upsampler_produces_correct_length() {
+        let mut us = Upsampler8to48::new();
+        let out = us.process(&vec![0i16; 160]);
+        assert_eq!(out.len(), 960);
+        let out2 = us.process(&vec![0i16; 160]);
+        assert_eq!(out2.len(), 960);
+    }
+
+    #[test]
+    fn fir_kernel_has_unity_dc_gain() {
+        let kernel = build_fir_kernel();
+        let sum: f64 = kernel.iter().sum();
+        assert!(
+            (sum - 1.0).abs() < 1e-10,
+            "FIR kernel DC gain should be 1.0, got {sum}"
+        );
+    }
+
+    #[test]
+    fn bessel_i0_known_values() {
+        // I₀(0) = 1
+        assert!((bessel_i0(0.0) - 1.0).abs() < 1e-12);
+        // I₀(1) ≈ 1.2660658
+        assert!((bessel_i0(1.0) - 1.2660658).abs() < 1e-5);
+    }
 }