From: http://support.apple.com/kb/TA40730

This note is about the cassette interface built into the Apple II and
Apple II+, subroutines. An assumption made here is that the cassette
recorder is in the proper mode, play or record, when the read and write
routines are executed. Note also that the timing is approximate and may
vary from one Apple to another.

A record is a block of binary data. This data may be a BASIC or APPLESOFT
program, a machine language program, or just binary data. Records representing
BASIC or APPLESOFT programs are really two records, the length of the program
and the actual program. A record consists of a header, synchronous bit, the
actual data, and a checksum byte for error detection.

Monitor record format

+--------+-+-----------------------+-+
| HEADER |S| DATA                  |C|
+--------+-+-----------------------+-+

BASIC program record format

+--------+-+----+-+--------+-+----------------+-+
| HEADER |S| LB |C| HEADER |S| PROGRAM        |C|
+--------+-+----+-+--------+-+----------------+-+

Key: S = SYNC bit
C = CHECKSUM byte
LB = BASIC program length

The header consists of 10 seconds of 770 Hz tone, (1 cycle equals 1300
microseconds). This gives enough time for the cassette motor to attain speed
and the plastic tape leader to go by. A subroutine called HEADR generates a
shortened header between the BASIC length bytes and the BASIC program itself.
The length of the header tone is controlled by the value of the accumulator on
entry to the subroutine. This can vary from 0.2 seconds to 40 seconds. On
entry the X register should be 0 and the carry flag should be set. HEADR also
generates a synchronous bit at the end of the tone. HEADR resides at
hexadecimal address $FCC9, or decimal address -882.

The last cycle of header tone and SYNC bit

---+             +-------------+    +-----+
   |             |             |    |     |
   +-------------+             +----+     |

| 1300 microseconds | 200 | 250 | header tone | synchronous bit |

The synchronous bit, generated by HEADR, is one half cycle of 2500 Hz, (200
microseconds) and one half cycle of 2000 Hz, (250 microseconds). It is used to
signal the end of the header tone and the start of the data.

The data is recorded on the tape with a low starting address and a high ending
address. Each byte of data is shifted out most significant bit first, least
significant bit last. A zero bit is made up of one cycle of 2 kHz, (250
microseconds per half cycle) and a one bit is one cycle of 1 kHz, (500
microseconds per half cycle). This works out to 2000 baud for zeros only and
1000 baud for ones, or an average of 1500 baud.

A zero bit and a one bit

+-----+     +----------+          +
|     |     |          |          |
+     +-----+          +----------+

| 500 usec | 1000 usec |

The checksum byte is written on the tape at the end of the data block. All
during reading or writing each data byte is EXCLUSIVE OR-ed with the checksum
byte. If the checksum computed during a read agrees with the checksum that was
written out, then the data is probably good. This method will detect an odd
number of errors for any of the eight bits of the byte.

In writing data, the cassette output uses quite simple circuitry, a flip-flop
connected through a voltage divider to the jack on the back panel of the
Apple. Any time the address $C020 is accessed this flip-flop changes state.
Accessing the flip-flop once every 500 microseconds generates a 1000 Hz
tone.