diff --git a/src/conversions/sample_rate/rubato.rs b/src/conversions/sample_rate/rubato.rs
index df1ee016..b321b640 100644
--- a/src/conversions/sample_rate/rubato.rs
+++ b/src/conversions/sample_rate/rubato.rs
@@ -193,19 +193,41 @@ impl<I: Source, R: rubato::Resampler<Sample>> RubatoResample<I, R> {
     }
 
     fn fill_input_buffer(&mut self, needed_by_resampler: InFrameCount) -> InFrameCount {
-        let current_span_length = self.input.current_span_len().map(InSamples);
-        let frames_to_take = needed_by_resampler
-            .samples(self.output.channels)
-            .min(current_span_length.unwrap_or(InSamples::MAX));
-
+        let wanted = needed_by_resampler.samples(self.output.channels);
         self.input_buffer.clear();
-        for _ in 0..frames_to_take.raw() {
-            if let Some(sample) = self.input.next() {
-                self.input_buffer.push(sample);
-            } else {
+
+        // Keep pulling across span boundaries while the format is unchanged.
+        // A span boundary only matters to the resampler when the sample rate
+        // or channel count actually changes; ending a chunk at every boundary
+        // would flag it `partial_len`, which rubato zero-pads.
+        while self.input_buffer.len() < wanted {
+            let span = self.input.current_span_len().map(InSamples);
+            let take = (wanted - self.input_buffer.len()).min(span.unwrap_or(InSamples::MAX));
+            if take == InSamples::ZERO {
+                break;
+            }
+
+            let rate_before = self.input.sample_rate();
+            let channels_before = self.input.channels();
+
+            let mut taken = 0usize;
+            for _ in 0..take.raw() {
+                if let Some(sample) = self.input.next() {
+                    self.input_buffer.push(sample);
+                    taken += 1;
+                } else {
+                    break;
+                }
+            }
+
+            if taken == 0
+                || self.input.sample_rate() != rate_before
+                || self.input.channels() != channels_before
+            {
                 break;
             }
         }
+
         self.input_buffer.len().frames(self.output.channels)
     }
 
diff --git a/src/conversions/sample_rate/tests.rs b/src/conversions/sample_rate/tests.rs
index c2ea52cc..5412ee9f 100644
--- a/src/conversions/sample_rate/tests.rs
+++ b/src/conversions/sample_rate/tests.rs
@@ -338,3 +338,67 @@ fn test_span_boundary_same_format() {
         output.len()
     );
 }
+
+/// Resampling must not depend on how the input is chunked into spans.
+///
+/// Decoders report one span per decoded packet (Vorbis packets are commonly
+/// 128–2048 frames, switching to short blocks around transients), so a single
+/// continuous stream arrives as many small spans. If a span boundary ends a
+/// resampler chunk, the chunk is flagged `partial_len` and zero-padded,
+/// injecting a discontinuity at every boundary.
+#[test]
+fn small_spans_do_not_degrade_resampled_signal() {
+    let source_rate = SampleRate::new(48000).unwrap();
+    let target_rate = SampleRate::new(44100).unwrap();
+    let channels = ChannelCount::new(2).unwrap();
+
+    // 0.5 s of a 440 Hz stereo sine at 48 kHz.
+    let signal: Vec<Sample> = (0..24_000)
+        .flat_map(|i| {
+            let t = i as Sample / source_rate.get() as Sample;
+            let s = (std::f64::consts::TAU as Sample * 440.0 * t).sin() * 0.5;
+            [s, s]
+        })
+        .collect();
+
+    // Fraction of output energy sitting at the 440 Hz fundamental, via
+    // Goertzel on the left channel. A clean resample keeps this ~1.0.
+    let tone_energy_fraction = |samples: &[Sample]| -> f64 {
+        // Accumulate in f64 regardless of `Sample`: the cast is a no-op
+        // under the `64bit` feature but needed without it.
+        #[allow(clippy::unnecessary_cast)]
+        let left: Vec<f64> = samples.iter().step_by(2).map(|&s| s as f64).collect();
+        let n = left.len();
+        let bin = 440.0 * n as f64 / target_rate.get() as f64;
+        let w = std::f64::consts::TAU * bin / n as f64;
+        let (mut prev, mut prev2) = (0.0, 0.0);
+        for &s in &left {
+            let current = s + 2.0 * w.cos() * prev - prev2;
+            prev2 = prev;
+            prev = current;
+        }
+        let tone = prev * prev + prev2 * prev2 - 2.0 * w.cos() * prev * prev2;
+        let total: f64 = left.iter().map(|s| s * s).sum();
+        (tone / (n as f64 / 2.0)) / total.max(f64::EPSILON)
+    };
+
+    for frames_per_span in [4096usize, 2048, 1024, 512, 256, 128] {
+        let span = frames_per_span * channels.get() as usize;
+        let mut chunks = signal.chunks(span);
+        let mut source = TestSource::new(chunks.next().unwrap().to_vec(), source_rate, channels);
+        for chunk in chunks {
+            source = source.chain(chunk.to_vec(), source_rate, channels);
+        }
+
+        let output: Vec<Sample> =
+            SampleRateConverter::new(source, target_rate, ResampleConfig::poly().build()).collect();
+        let fraction = tone_energy_fraction(&output);
+
+        assert!(
+            fraction > 0.99,
+            "{frames_per_span}-frame spans: only {:.1}% of output energy at the \
+             fundamental (expected >99%) — span boundaries are corrupting the signal",
+            fraction * 100.0
+        );
+    }
+}