diff --git a/README.md b/README.md
index 25ccaef..c61312f 100644
--- a/README.md
+++ b/README.md
@@ -204,7 +204,7 @@ Here's the full list:
 | `l`    | `vel` | float 0-1+ | Velocity: > 0 to trigger note on, 0 to trigger note off |
 | `L`    | `mod_source` | 0 to OSCS-1 | Which oscillator is used as an modulation/LFO source for this oscillator. Source oscillator will be silent. |
 | `m`    | `portamento` | uint | Time constant (in ms) for pitch changes when note is changed without intervening note-off.  default 0 (immediate), 100 is good. |
-| `M`    | `echo` | float[,int,int,float,float] | Echo parameters --  level, delay_ms, max_delay_ms, feedback, 1st order IIR zero location. |
+| `M`    | `echo` | float[,int,int,float,float] | Echo parameters --  level, delay_ms, max_delay_ms, feedback, filter_coef (-1 is HPF, 0 is flat, +1 is LPF). |
 | `N`    | `note` | uint 0-127 | Midi note, sets frequency |
 | `N`    | `latency_ms` | uint | Sets latency in ms. default 0 (see LATENCY) |
 | `o`    | `algorithm` | uint 1-32 | DX7 FM algorithm to use for ALGO type |
diff --git a/src/amy.c b/src/amy.c
index f2e28d1..1366bee 100644
--- a/src/amy.c
+++ b/src/amy.c
@@ -181,17 +181,20 @@ void config_echo(float level, float delay_ms, float max_delay_ms, float feedback
             echo.max_delay_samples = max_delay_samples;
             //fprintf(stderr, "config_echo: max_delay_samples=%d\n", max_delay_samples);
         }
-        // Apply delay.
-        if (delay_samples > echo.max_delay_samples) delay_samples = echo.max_delay_samples;
+        // Apply delay.  We have to stay 1 sample less than delay line length for FIR EQ delay.
+        if (delay_samples > echo.max_delay_samples - 1) delay_samples = echo.max_delay_samples - 1;
         for (int c = 0; c < AMY_NCHANS; ++c) {
             echo_delay_lines[c]->fixed_delay = delay_samples;
         }
     }
     echo.level = F2S(level);
     echo.delay_samples = delay_samples;
-    echo.feedback = F2S(feedback);
-    // FIR filter is [1], [1, filter_coef] but scaled to have peak gain of 1.0.
+    // Filter is IIR [1, filter_coef] normalized for filter_coef > 0 (LPF), or FIR [1, filter_coef] normalized for filter_coef < 0 (HPF).
+    if (filter_coef > 0.99)  filter_coef = 0.99;  // Avoid unstable filters.
     echo.filter_coef = F2S(filter_coef);
+    // FIR filter potentially has gain > 1 for high frequencies, so discount the loop feedback to stop things exploding.
+    if (filter_coef < 0)  feedback /= 1.f - filter_coef;
+    echo.feedback = F2S(feedback);
     //fprintf(stderr, "config_echo: delay_samples=%d level=%.3f feedback=%.3f filter_coef=%.3f fc0=%.3f\n", delay_samples, level, feedback, filter_coef, S2F(echo.filter_coef));
 }
 
diff --git a/src/amy_config.h b/src/amy_config.h
index 8e3aa68..bfdadc5 100644
--- a/src/amy_config.h
+++ b/src/amy_config.h
@@ -98,7 +98,8 @@ extern const uint16_t pcm_samples;
 // echo setup,  Levels etc are SAMPLE (fxpoint), delays are in samples.
 #define ECHO_DEFAULT_LEVEL 0
 #define ECHO_DEFAULT_DELAY_MS  500.f
-#define ECHO_DEFAULT_MAX_DELAY_MS 1000.f
+// Delay line allocates in 2^n samples at 44k; 743ms is just under 32768 samples.
+#define ECHO_DEFAULT_MAX_DELAY_MS 743.f
 #define ECHO_DEFAULT_FEEDBACK 0
 #define ECHO_DEFAULT_FILTER_COEF 0
 
diff --git a/src/delay.c b/src/delay.c
index 070c765..d5a3c35 100644
--- a/src/delay.c
+++ b/src/delay.c
@@ -85,9 +85,9 @@ void delay_line_in_out(SAMPLE *in, SAMPLE *out, int n_samples, SAMPLE* mod_in, S
     delay_line->next_in = index_in;
 }
 
-static inline SAMPLE DEL_OUT(delay_line_t *delay_line, int offset) {
+static inline SAMPLE DEL_OUT(delay_line_t *delay_line, int extra_delay) {
     int out_index =
-        (delay_line->next_in - (delay_line->fixed_delay - offset)) & (delay_line->len - 1);
+        (delay_line->next_in - (delay_line->fixed_delay + extra_delay)) & (delay_line->len - 1);
     return delay_line->samples[out_index];
 }
 
@@ -108,17 +108,36 @@ void delay_line_in_out_fixed_delay(SAMPLE *in, SAMPLE *out, int n_samples, int d
     // Read and write the next n_samples from/to the delay line.
     // Simplified version of delay_line_in_out() that uses a fixed integer delay
     // for the whole block.
-    while(n_samples-- > 0) {
-        SAMPLE next_in = *in++;
-        SAMPLE last_out = delay_line->samples[(delay_line->next_in - 1) & (delay_line->len - 1)];
-        //SAMPLE sample_1 = DEL_OUT(delay_line, 1);
-        SAMPLE sample = DEL_OUT(delay_line, 0);
-        if (filter_coef > 0)
-            sample += SMULR6(filter_coef, last_out - sample);
-        else if (filter_coef < 0)
-            sample += SMULR6(filter_coef, last_out + sample);
-        DEL_IN(delay_line, next_in + SMULR6(feedback_level, sample));
-        *out++ += MUL8_SS(mix_level, sample);  // mix delayed + original.
+    if (filter_coef == 0) {
+        while(n_samples-- > 0) {
+            SAMPLE delay_out = DEL_OUT(delay_line, 0);
+            SAMPLE next_in = *in++ + SMULR6(feedback_level, delay_out);
+            DEL_IN(delay_line, next_in);
+            *out++ += MUL8_SS(mix_level, delay_out);
+        }
+    } else if (filter_coef > 0) {
+        // Positive filter coef is a pole on the positive real line, to get low-pass effect.
+        // We apply the filter on the way *in* to the delay line.
+        while(n_samples-- > 0) {
+            SAMPLE delay_out = DEL_OUT(delay_line, 0);
+            SAMPLE next_in = *in++ + SMULR6(feedback_level, delay_out);
+            // Peek at the last value we wrote to the delay line to get the most recent filter output.
+            SAMPLE last_filter_result = delay_line->samples[(delay_line->next_in - 1) & (delay_line->len - 1)];
+            SAMPLE filter_result = next_in + SMULR6(filter_coef, last_filter_result - next_in);
+            DEL_IN(delay_line, filter_result);
+            *out++ += MUL8_SS(mix_level, delay_out);
+        }
+    } else {
+        // Negative filter coef is a zero on the positive real axis to get high-pass.
+        // We apply the FIR zero on the way *out* of the delay line.
+        while(n_samples-- > 0) {
+            SAMPLE delay_out = DEL_OUT(delay_line, 0);
+            SAMPLE prev_delay_out = DEL_OUT(delay_line, 1);
+            SAMPLE output = delay_out + SMULR6(filter_coef, prev_delay_out);
+            SAMPLE next_in = *in++ + SMULR6(feedback_level, output);
+            DEL_IN(delay_line, next_in);
+            *out++ += MUL8_SS(mix_level, output);
+        }
     }
 }
 
diff --git a/test.py b/test.py
index 25cf9ad..f963680 100644
--- a/test.py
+++ b/test.py
@@ -467,7 +467,25 @@ def run(self):
     amy.send(time=0, osc=0, bp0="0,1,200,0,0,0")
 
     amy.send(time=100, osc=0, note=48, vel=1)
-    
+
+
+class TestEchoLPF(AmyTest):
+
+  def run(self):
+    amy.echo(level=0.5, delay_ms=200, feedback=0.7, filter_coef=0.9)
+    amy.send(time=0, osc=0, wave=amy.SAW_DOWN, bp0="0,1,200,0,0,0")
+
+    amy.send(time=100, osc=0, note=48, vel=1)
+
+
+class TestEchoHPF(AmyTest):
+
+  def run(self):
+    amy.echo(level=0.5, delay_ms=200, feedback=0.7, filter_coef=-0.9)
+    amy.send(time=0, osc=0, wave=amy.SAW_DOWN, bp0="0,1,200,0,0,0")
+
+    amy.send(time=100, osc=0, note=48, vel=1)
+
 
 def main(argv):
   if len(argv) > 1:
diff --git a/tests/ref/TestEcho.wav b/tests/ref/TestEcho.wav
index e8d2984..f192afb 100644
Binary files a/tests/ref/TestEcho.wav and b/tests/ref/TestEcho.wav differ
diff --git a/tests/ref/TestEchoHPF.wav b/tests/ref/TestEchoHPF.wav
new file mode 100644
index 0000000..81cf3f4
Binary files /dev/null and b/tests/ref/TestEchoHPF.wav differ
diff --git a/tests/ref/TestEchoLPF.wav b/tests/ref/TestEchoLPF.wav
new file mode 100644
index 0000000..0a137a9
Binary files /dev/null and b/tests/ref/TestEchoLPF.wav differ