diff --git a/apps/casette/encode.py b/apps/casette/encode.py
index 352c001..55a4d37 100644
--- a/apps/casette/encode.py
+++ b/apps/casette/encode.py
@@ -1,133 +1,135 @@
 #!/usr/bin/env python3
 from __future__ import annotations
 
-import os
 import argparse
-import math
 import struct
 import wave
 from pathlib import Path
 
-class FSKWriter:
-    def __init__(self, sample_rate: int, c_low: int, c_high: int, f_low: float, f_high: float, amplitude: float, block_spacing: float):
+
+class PulseWriter:
+    def __init__(
+        self,
+        sample_rate: int,
+        t_short_us: float,
+        t_long_us: float,
+        amplitude: float,
+    ) -> None:
         self.sample_rate = sample_rate
-        self.c_low = c_low
-        self.c_high = c_high
-        self.f_low = f_low
-        self.f_high = f_high
-        self.amplitude = amplitude
-        self.samples_per_low = (sample_rate * c_low) // f_low
-        self.samples_per_high = (sample_rate * c_high) // f_high
-        self.phase = 0.0
-        self.block_spacing = block_spacing
+        self.short_samples = self._duration_samples(t_short_us)
+        self.long_samples = self._duration_samples(t_long_us)
+        self.high = round(32767 * amplitude)
+        self.low = -self.high
         self.samples = bytearray()
 
-        self.space(1.0)
+        self.space(0.1)
 
-    def finalize(self):
-        self.space(1.0)
+    def space(self, time: float) -> None:
+        self._level(0, int(time*self.sample_rate))
 
-    def space(self, time:float = 0.03):
-        self.samples.extend(b'\0\0'*int(time*self.sample_rate))
+    def _duration_samples(self, duration_us: float) -> int:
+        sample_count = round(duration_us * self.sample_rate / 1_000_000)
+        if sample_count < 1:
+            raise ValueError(
+                f"pulse duration {duration_us:g} us is shorter than one sample"
+            )
+        return sample_count
 
-    def _tone(self, frequency: float, sample_count: int) -> None:
-        phase_step = 2.0 * math.pi * frequency / self.sample_rate
-        scale = int(32767 * self.amplitude)
-        phase = self.phase
-        out = self.samples
+    def _level(self, value: int, sample_count: int) -> None:
+        self.samples.extend(struct.pack("<h", value) * sample_count)
 
-        for _ in range(int(sample_count)):
-            sample = int(round(scale * math.sin(phase)))
-            out.extend(struct.pack("<h", sample))
-            phase += phase_step
-            if phase >= 2.0 * math.pi:
-                phase -= 2.0 * math.pi
+    def short_pulse(self) -> None:
+        self._level(self.high, self.short_samples)
+        self._level(self.low, self.long_samples)
 
-        self.phase = phase
+    def long_pulse(self) -> None:
+        self._level(self.high, self.long_samples)
+        self._level(self.low, self.short_samples)
 
-    def leader(self) -> None:
-        self._tone(self.f_high, (200*self.sample_rate)//self.f_high)
-        self._tone(self.f_low, (200*self.sample_rate)//self.f_low)
-        for _ in range(8):
-            self.bit(1)
-            self.bit(0)
-        self.byte(0xf0)
+    def marker(self) -> None:
+        self._level(self.high, 2*self.long_samples)
+        self._level(self.low, 2*self.long_samples)
 
     def bit(self, value: int) -> None:
         if value:
-            self._tone(self.f_high, self.samples_per_high)
+            self.long_pulse()
         else:
-            self._tone(self.f_low, self.samples_per_low)
+            self.short_pulse()
 
     def byte(self, value: int) -> None:
-        # MSB first
-        for bit_no in range(7, -1, -1):
-            self.bit((value >> bit_no) & 1)
+        # Transmit bytes most-significant bit first.
+        for bit_number in range(7, -1, -1):
+            self.bit((value >> bit_number) & 1)
 
-    def bytes(self, data: bytes) -> None:
-        for value in data:
+    def bytes(self, value: bytes) -> None:
+        for byte in value:
+            self.byte(byte)
+
+    def leader(self) -> None:
+        self.space(0.2)
+        for _ in range(50):
+            self._level(self.high, self.short_samples)
+            self._level(self.low, self.short_samples)
+        for _ in range(50):
+            self._level(self.high, self.long_samples)
+            self._level(self.low, self.long_samples)
+        self.marker()
+        for value in (0x55, 0x5A, 0xC3):
             self.byte(value)
 
-    def block(self, nr: int, data: bytes) -> None:
-        print(f'> {data}') 
-        self.leader()
-        self.byte(nr)
-        l = len(data)
-        self.byte(l>>8)
-        self.byte(l&0xff)
-        #self.samples.extend(b'\0'*int(self.samples_per_low*8))
-        self.bytes(data)
-        self.byte(self.chsum(data))
-        self.space(self.block_spacing)
-
-    def chsum(self, data: bytes) -> int:
-        cs = 0
-        for d in data:
-            cs += int(d)
-        return cs
-
 
 def main() -> None:
-    parser = argparse.ArgumentParser(description="Encode a binary file as two tone FSK WAV")
-    parser.add_argument("input", type=Path, help="Binary input file")
-    parser.add_argument("output", type=Path, help="output WAV; default <input>.wav")
-    parser.add_argument("--sample-rate", type=int, default=48000)
-    parser.add_argument("--low", type=float, default=2000.0, help="low tone in Hz")
-    parser.add_argument("--high", type=float, default=4000.0, help="high tone in Hz")
-    parser.add_argument("--low-cycles", type=int, default=2, help="low tone cycle count for a 0")
-    parser.add_argument("--high-cycles", type=int, default=4, help="high tone cycle count for a 1")
+    parser = argparse.ArgumentParser(
+        description="Generate the pulse-width encoded tape leader"
+    )
+    parser.add_argument("output", type=Path, help="output WAV file")
+    parser.add_argument("--sample-rate", type=int, default=48_000)
+    parser.add_argument(
+        "--t-short",
+        type=float,
+        default=250.0,
+        metavar="MICROSECONDS",
+        help="short pulse interval in microseconds (default: 250)",
+    )
+    parser.add_argument(
+        "--t-long",
+        type=float,
+        default=500.0,
+        metavar="MICROSECONDS",
+        help="long pulse interval in microseconds (default: 500)",
+    )
     parser.add_argument("--amplitude", type=float, default=0.75)
-    parser.add_argument("--block-size", type=int, default=128)
-    parser.add_argument("--block-spacing", type=float, default=1.0)
     args = parser.parse_args()
 
-    output_path = args.output or args.input.with_name(args.input.stem + ".wav")
+    if args.sample_rate <= 0:
+        parser.error("--sample-rate must be positive")
+    if args.t_short <= 0 or args.t_long <= 0:
+        parser.error("pulse intervals must be positive")
+    if args.t_short >= args.t_long:
+        parser.error("--t-short must be less than --t-long")
+    if not 0.0 <= args.amplitude <= 1.0:
+        parser.error("--amplitude must be between 0 and 1")
 
-    tones = FSKWriter(args.sample_rate, args.low_cycles, args.high_cycles, args.low, args.high, args.amplitude, args.block_spacing)
+    try:
+        pulses = PulseWriter(
+            args.sample_rate,
+            args.t_short,
+            args.t_long,
+            args.amplitude,
+        )
+    except ValueError as error:
+        parser.error(str(error))
 
-    # Create header block
-    nrblocks = math.ceil(os.path.getsize(args.input)/args.block_size)
-    print(f"Number of blocks {nrblocks}*2")
+    pulses.leader()
+    pulses.bytes(b'Hello world!\0')
 
-    tones.block(0, struct.pack('<H', nrblocks*2) + args.input.name.encode() + b'\0')
-    tones.space(1.0)
-
-    block = 1
-    with open(args.input, 'rb') as f:
-        while True:
-            data = f.read(args.block_size)
-            if not len(data):
-                break
-            tones.block(block, data)
-            tones.block(block, data)
-            block += 1
-
-    tones.finalize()
-    with wave.open(str(output_path), "wb") as wav:
+    pulses.space(1.0)
+    with wave.open(str(args.output), "wb") as wav:
         wav.setnchannels(1)
         wav.setsampwidth(2)
         wav.setframerate(args.sample_rate)
-        wav.writeframes(tones.samples)
+        wav.writeframes(pulses.samples)
+
 
 if __name__ == "__main__":
     main()
diff --git a/apps/casette/tape_loader.asm b/apps/casette/tape_loader.asm
index 975b26c..eecb3ba 100644
--- a/apps/casette/tape_loader.asm
+++ b/apps/casette/tape_loader.asm
@@ -3,6 +3,21 @@
 .area   TAPELOAD (ABS)
 .org    TAPE_LOADER_ROM_ADDRESS
 
+.macro print_var arg1
+	ld		a, (arg1)
+	ei
+	call	print_hex8
+	call	print_crlf
+	di
+.endm
+
+.macro error_ret arg1,arg2
+arg1:
+	ld		a,#arg2
+	scf
+	ret
+.endm
+
 TAPE_PORT       = 0x18
 MAX_BLOCK_SIZE  = 0x100
 
@@ -11,6 +26,12 @@ WORK_BASE       = 0xB000
 
 OLD_SP          = WORK_BASE + 0x00 ; word
 OLD_IFF         = WORK_BASE + 0x02 ; byte
+SHORT_AVG		= WORK_BASE + 0x03 ; byte
+SHORT_MAX		= WORK_BASE + 0x04 ; byte
+SHORT_MIN		= WORK_BASE + 0x05 ; byte
+LONG_AVG		= WORK_BASE + 0x06 ; byte
+LONG_MAX		= WORK_BASE + 0x07 ; byte
+LONG_MIN		= WORK_BASE + 0x08 ; byte
 
 BLOCK_BUFFER    = WORK_BASE + 0x100 ; MAX_BLOCK_SIZE
 
@@ -36,6 +57,52 @@ main:
     ld      hl,#msg_start
     call    print_string_ei
 
+	; Calibrate reader
+	call	calibrate
+	jr		nc,1$
+	ei
+	call	print_hex8
+	call	print_crlf	
+	ld		hl,#msg_cal_fail
+	call	print_string
+	di
+	jp		exit
+1$:
+	; try to read the sync bytes
+	call	read_sync
+	jr		nc,2$
+	ld		hl,#msg_sync_error
+	call	print_string_ei
+	jp		exit
+2$:
+
+	; Test, read Hello World!\0 (13 bytes)
+	exx
+	ld		e,#0
+	ld		hl,#BLOCK_BUFFER
+	exx
+4$:
+	call	read_byte
+	jr		nc,5$
+	ld		hl,#msg_read_error
+	call	print_string_ei
+	exx
+	ei
+	call	print_hex16
+	call	print_crlf
+	di
+	exx
+	jp		exit
+5$:
+	exx
+	ld		(hl),a
+	inc		hl
+	inc		e
+	ld		a,e
+	cp		a,#13
+	exx
+	jr		nz,4$
+
 ; #################################################3
 ; Exit
 ; #################################################3
@@ -57,6 +124,379 @@ exit:
 return_di:
     ret
 
+; #################################################3
+; Measure edge-to-edge pulse
+; needs d (comparator state)
+; returns b (loop count)
+; carry is timeout
+; uses a
+; loop is 36T, extra 22T (0.6 loops)
+; #################################################3
+measure_pulse:
+	ld		b,#0			; 4T
+1$:
+	; Use b as timeout counter (256 loops)
+	inc		b				; 4T
+	jr		z,9$			; 12/7T
+	in		a,(TAPE_PORT)	; 11T
+	cp		a,d				; 4T
+	jp		z,1$			; 10T
+	; loop has 36T
+	; Edge found
+	ld		d,a				; 4T
+	or		a,a				; 4T
+	ret						; 10T
+9$:
+	scf
+	ret
+
+; #################################################3
+; Ignore one pulse then measure next pulse
+; needs d (comparator state)
+; returns b (loop count)
+; carry is timeout
+; uses a
+; loop is 36T, extra 26T (0.7 loops)
+; #################################################3
+sample_pulse:
+	; wait for first edge
+	ld		b,#0			; 4T
+1$:
+	; Use b as timeout counter (256 loops)
+	inc		b				; 4T
+	jr		z,9$			; 12/7T
+	in		a,(TAPE_PORT)	; 11T
+	cp		a,d				; 4T
+	jp		z,1$			; 10T
+	; loop has 36T
+	; Edge found
+	ld		d,a				; 4T
+	; start timing pulse length
+	ld		b,#0			; 4T
+2$:
+	; Use b as timeout counter (256 loops)
+	inc		b				; 4T
+	jr		z,9$			; 12/7T
+	in		a,(TAPE_PORT)	; 11T
+	cp		a,d				; 4T
+	jp		z,2$			; 10T
+	; loop has 36T
+	; Edge found
+	ld		d,a				; 4T
+	or		a,a				; 4T
+	ret						; 10T
+9$:
+	scf
+	ret
+
+; #################################################3
+; Average 8 pulse measurements
+; needs d (comparator state)
+; returns a (average)
+; carry is timeout
+; uses b,e,h
+; #################################################3
+average8:
+	ld		e,#0
+	ld		h,#8
+1$:
+	; synchronize to next pulse
+	; ignore first measurement
+	; contains part of the calculations
+	call	sample_pulse
+	call	sample_pulse
+	; when timeout just return with carry
+	ret		c
+	; Add loop count to e
+	ld		a,e
+	add		a,b
+	ld		e,a
+	; repeat 8 times
+	dec		h
+	jr		nz,1$
+	; divide by 8
+	ld		a,e
+	srl		a
+	srl		a
+	srl		a
+	; clear carry
+	or		a,a
+	ret
+
+; #################################################3
+; Make +/-25% window around a
+; needs a (average)
+; returns b (minimum) c (maximum)
+; uses e,h
+; NOTE: USES 8 BIT COUNTER
+;       @1MHZ t_long<1.14ms
+;       @2MHZ t_long<571us
+;       @4MHZ t_long<285us
+; TODO: Check feasibility of using 16bit accumulator
+; #################################################3
+make_window:
+	ld      e,a
+	; get 25%
+	srl     a
+	srl     a
+	ld      h,a
+	; a-25%
+	ld      a,e
+	sub     h
+	ld      b,a
+	; a+25%
+	ld      a,e
+	add     a,h
+	ld      c,a
+	ret
+
+; #################################################3
+; Calibration
+; #################################################3
+calibrate:
+	; Get current comparator state
+	in		a,(TAPE_PORT)
+	ld		d,a
+
+	; make sure we start in silence
+	call	sample_pulse
+	; should timeout
+	jp		nc,cal_fail_no_silence
+
+	; Wait for first edge
+	ld		l,#0
+1$:
+	call	sample_pulse
+	jr		nc,2$
+	; Use b as timeout counter (256 loops)
+	; inc		l
+	; jr		nz,1$
+	jr		1$
+	; timeout
+	;jp		cal_fail_no_short_leader ; TODO IGNORED TIMEOUT FOR NOW HERE
+2$:
+	; edge found
+
+	; get short leader timing
+	call	average8
+	jr		c,cal_fail_short
+	ld		(SHORT_AVG),a
+
+	call	make_window
+	ld		a,b
+	ld		(SHORT_MIN),a
+	ld		a,c
+	ld		(SHORT_MAX),a
+
+	; find long leader
+	; need 4 consecutive pulses > short*1.5
+	ld		a,(SHORT_AVG)
+	ld		c,a
+	srl		a
+	add		a,c
+	ld		c,a
+	; c = short*1.5
+	
+	; TODO no overal timeout here. a continuous 
+	; short pulse leader will keep on waiting 
+	; forever
+3$:
+	ld		e,#4
+4$:
+	; ignore first to get accurate measurement
+	call	sample_pulse
+	call	sample_pulse
+	jr		c,cal_fail_find_long
+	ld		a,b
+	; if it is no long pulse try again
+	cp		a,c
+	jr		c,3$
+	; long pulse found
+	dec     e
+	jr		nz,4$
+	; 4 (or well... 8) long pulses found
+
+	; get long leader timning
+	call	average8
+	jr		c,cal_fail_long
+	ld		(LONG_AVG),a
+
+	call	make_window
+	ld		a,b
+	ld		(LONG_MIN),a
+	ld		a,c
+	ld		(LONG_MAX),a
+
+	; A quick sanity check
+	; long > 1.5*short
+	ld		a,(SHORT_AVG)
+	ld		c,a
+	srl		a
+	add		a,c
+	ld		c,a
+	; c = 1.5*short
+	ld		a,(LONG_AVG)
+	cp		a,c
+	jr		c,cal_fail_sanity
+
+	; Find the marker
+	; marker is 2*long high 2*long low
+	; threshold is 1.5*long
+	ld		a,(LONG_AVG)
+	ld		c,a
+	srl		a
+	add		a,c
+	ld		c,a
+	; c = 1.5*long
+
+    ; Find two consecutive marker-width intervals.
+5$:
+    ld      e,#0
+6$:
+    call    measure_pulse
+    jr      c,cal_fail
+    ld      a,b
+    cp      a,c
+    jr      c,5$
+    inc     e
+    ld      a,e
+    cp      a,#2
+    jr      c,6$
+    ; Both marker halves consumed. We are now at the
+    ; beginning of the first data pulse.
+
+	; next pulse is first pulse of training byte
+	; must decode as 0x55
+	call	read_byte
+	jr		c,cal_fail_read_error
+	cp		a,#0x55
+	jr		nz,cal_fail_training_byte
+
+	ld		a,#0
+	or		a,a
+	ret
+
+error_ret cal_fail 0
+error_ret cal_fail_short 1
+error_ret cal_fail_find_long 2
+error_ret cal_fail_long 3
+error_ret cal_fail_sanity 4
+error_ret cal_fail_training_byte 5
+error_ret cal_fail_no_silence 6
+error_ret cal_fail_no_short_leader 7
+error_ret cal_fail_read_error 8
+
+; #################################################3
+; Read single bit
+; needs d (comparator state)
+; valid: z=1 c=bit
+; error: z=0
+; uses a,b,c
+; #################################################3
+read_bit:
+	; entry may be late into first pulse
+	; wait for first edge to measure
+	ld		b,#0
+1$:
+	inc		b
+	jr		z,bit_bad
+	in		a,(TAPE_PORT)
+	cp		a,d
+	jp		z,1$
+	ld		d,a
+	; ignored first edge
+	; measure second pulse
+	ld		b,#0
+2$:
+	inc		b
+	jr		z,bit_bad
+	in		a,(TAPE_PORT)
+	cp		a,d
+	jp		z,2$
+	ld		d,a
+	; if in short window (second half is short pulse)
+	; 1
+	ld		a,(SHORT_MIN)
+	ld		c,a
+	ld		a,b
+	cp		a,c
+	jr		c,bit_bad ; pulse<SHORT_MIN
+	ld		a,(SHORT_MAX)
+	ld		c,a
+	ld		a,b
+	cp		a,c
+	jr		c,bit_one
+	jr		z,bit_one ; pulse<=SHORT_MAX
+	; test long window (second half is long pulse)
+	ld		a,(LONG_MIN)
+	ld		c,a
+	ld		a,b
+	cp		a,c
+	jr		c,bit_bad ; pulse<LONG_MAX
+	ld		a,(LONG_MAX)
+	ld		c,a
+	ld		a,b
+	cp		a,c
+	jr		c,bit_zero
+	jr		z,bit_zero ; pulse<=LONG_MAX
+bit_bad:
+	ld		a,#1
+	or		a,a
+	ret
+bit_zero:
+	xor		a,a
+	ret
+bit_one:
+	xor		a,a
+	scf
+	ret
+
+; #################################################3
+; Read byte, MSB first
+; returns a (byte)
+; carry means error
+; needs d (comparator state)
+; uses e,h,b,c
+; #################################################3
+read_byte:
+	ld		e,#0
+	ld		h,#8
+1$:
+	call	read_bit
+	jr		nz,9$
+	; shift in the carry bit
+	rl		e
+	; repeat 8 times
+	dec		h
+	jr		nz,1$
+	; byte done, 
+	ld		a,e
+	or		a
+	ret
+
+9$:
+	scf
+	ret
+
+; #################################################3
+; Read and check sync bytes
+; carry is error
+; #################################################3
+read_sync:
+	call	read_byte
+	ret		c
+	cp		a,#0x5a
+	jr		nz,9$
+	call	read_byte
+	ret		c
+	cp		a,#0xc3
+	jr		nz,9$
+	or		a,a
+	ret
+9$:
+	scf
+	ret
+
 ; #################################################3
 ; IO Helper functions
 ; #################################################3
@@ -118,3 +558,12 @@ print_hex_digit:
 msg_start:
     .ascii  "TAPE DECODE TEST - PLAY TAPE"
     .db     13,10,0
+msg_cal_fail:
+	.ascii 	"Calibration failed"
+    .db     13,10,0
+msg_sync_error:
+	.ascii	"Sync bytes not found"
+	.db		13,10,0
+msg_read_error:
+	.ascii	"read error"
+	.db		13,10,0
\ No newline at end of file
diff --git a/apps/casette/test2.wav b/apps/casette/test2.wav
index 159e561..3958d66 100644
Binary files a/apps/casette/test2.wav and b/apps/casette/test2.wav differ
diff --git a/apps/casette/tone_2khz.wav b/apps/casette/tone_2khz.wav
index 48a57ff..7db8a2b 100644
Binary files a/apps/casette/tone_2khz.wav and b/apps/casette/tone_2khz.wav differ
diff --git a/apps/casette/tone_4khz.wav b/apps/casette/tone_4khz.wav
index 98bc5ab..a7276da 100644
Binary files a/apps/casette/tone_4khz.wav and b/apps/casette/tone_4khz.wav differ
diff --git a/apps/casette/tone_4khz_2khz.wav b/apps/casette/tone_4khz_2khz.wav
new file mode 100644
index 0000000..0754286
Binary files /dev/null and b/apps/casette/tone_4khz_2khz.wav differ