2020-08-11 22:26:55 +00:00
|
|
|
#!/usr/bin/env python3
|
2020-08-11 22:23:33 +00:00
|
|
|
# Delta modulation audio encoder.
|
|
|
|
#
|
2020-08-16 22:15:30 +00:00
|
|
|
# Simulates the Apple II speaker at 1MHz (i.e. cycle-level) resolution,
|
|
|
|
# by modeling it as an RC circuit with given time constant. In order to
|
|
|
|
# reproduce a target audio waveform, we upscale it to 1MHz sample rate,
|
|
|
|
# and compute the sequence of player opcodes to best reproduce this waveform.
|
2020-08-11 22:23:33 +00:00
|
|
|
#
|
2020-12-28 22:42:34 +00:00
|
|
|
# XXX
|
2020-08-16 22:15:30 +00:00
|
|
|
# Since the player opcodes are chosen to allow ticking the speaker during any
|
|
|
|
# given clock cycle (though with some limits on the minimum time
|
|
|
|
# between ticks), this means that we are able to control the Apple II speaker
|
|
|
|
# with cycle-level precision, which results in high audio fidelity with low
|
|
|
|
# noise.
|
|
|
|
#
|
|
|
|
# To further optimize the audio quality we look ahead some defined number of
|
|
|
|
# cycles and choose a speaker trajectory that minimizes errors over this range.
|
|
|
|
# e.g. this allows us to anticipate large amplitude changes by pre-moving
|
2020-08-11 22:23:33 +00:00
|
|
|
# the speaker to better approximate them.
|
|
|
|
#
|
2020-12-28 22:42:34 +00:00
|
|
|
# This also needs to take into account scheduling the "end of frame" opcode
|
|
|
|
# every 2048 output bytes, where the Apple II will manage the TCP socket buffer
|
|
|
|
# while ticking the speaker at a regular cadence to keep it in a net-neutral
|
|
|
|
# position. When looking ahead we can also (partially) compensate for this
|
|
|
|
# "dead" period by pre-positioning.
|
2020-08-11 22:23:33 +00:00
|
|
|
|
2020-12-28 12:54:44 +00:00
|
|
|
import argparse
|
2020-08-16 22:15:30 +00:00
|
|
|
import collections
|
2020-08-24 21:28:28 +00:00
|
|
|
import functools
|
2020-08-11 22:23:33 +00:00
|
|
|
import librosa
|
2020-08-10 20:03:12 +00:00
|
|
|
import numpy
|
2020-08-11 22:23:33 +00:00
|
|
|
from eta import ETA
|
2020-08-10 20:03:12 +00:00
|
|
|
|
2020-08-13 21:08:50 +00:00
|
|
|
import opcodes
|
2020-08-10 20:03:12 +00:00
|
|
|
|
2022-05-16 20:11:17 +00:00
|
|
|
import lookahead
|
|
|
|
|
2020-08-24 21:28:28 +00:00
|
|
|
|
2020-12-07 20:48:15 +00:00
|
|
|
# We simulate the speaker voltage trajectory resulting from applying multiple
|
|
|
|
# voltage profiles, compute the resulting squared error relative to the target
|
|
|
|
# waveform, and pick the best one.
|
|
|
|
#
|
|
|
|
# We use numpy to vectorize the computation since it has better scaling
|
|
|
|
# performance with more opcode choices, although also has a larger fixed
|
|
|
|
# overhead.
|
|
|
|
#
|
|
|
|
# The speaker position p_i evolves according to
|
|
|
|
# p_{i+1} = p_i + (v_i - p_i) / s
|
|
|
|
# where v_i is the i'th applied voltage, s is the speaker step size
|
|
|
|
#
|
|
|
|
# Rearranging, we get p_{i+1} = v_i / s + (1-1/s) p_i
|
|
|
|
# and if we expand the recurrence relation
|
|
|
|
# p_{i+1} = Sum_{j=0}^i (1-1/s)^(i-j) v_j / s + (1-1/s)^(i+1) p_0
|
|
|
|
# = (1-1/s)^(i+1)(1/s * Sum_{j=0}^i v_j / (1-1/s)^(j+1) + p0)
|
|
|
|
#
|
|
|
|
# We can precompute most of this expression:
|
|
|
|
# 1) the vector {(1-1/s)^i} ("_delta_powers")
|
|
|
|
# 2) the position-independent term of p_{i+1} ("_partial_positions"). Since
|
|
|
|
# the candidate opcodes list only depends on frame_offset, the voltage matrix
|
|
|
|
# v also only takes a few possible values, so we can precompute all values
|
|
|
|
# of this term.
|
2020-08-10 20:03:12 +00:00
|
|
|
|
2020-08-16 22:15:30 +00:00
|
|
|
|
2020-12-07 20:48:15 +00:00
|
|
|
@functools.lru_cache(None)
|
|
|
|
def _delta_powers(shape, step_size: int) -> numpy.ndarray:
|
|
|
|
delta = 1 - 1 / step_size
|
|
|
|
return numpy.cumprod(numpy.full(shape, delta), axis=-1)
|
2020-08-25 19:46:00 +00:00
|
|
|
|
2020-08-10 20:03:12 +00:00
|
|
|
|
2020-12-07 20:48:15 +00:00
|
|
|
def _partial_positions(voltages, step_size):
|
|
|
|
delta_powers = _delta_powers(voltages.shape, step_size)
|
2020-08-10 20:03:12 +00:00
|
|
|
|
2020-12-07 20:48:15 +00:00
|
|
|
partial_positions = delta_powers * (
|
|
|
|
numpy.cumsum(voltages / delta_powers, axis=-1) / step_size)
|
|
|
|
return delta_powers, partial_positions
|
2020-08-10 20:03:12 +00:00
|
|
|
|
2020-08-16 22:15:30 +00:00
|
|
|
|
2020-12-07 20:48:15 +00:00
|
|
|
def new_positions(
|
|
|
|
position: float, partial_positions: numpy.ndarray,
|
|
|
|
delta_powers: numpy.ndarray) -> numpy.ndarray:
|
|
|
|
"""Computes new array of speaker positions for position and voltage data."""
|
|
|
|
return partial_positions + delta_powers * position
|
2020-08-24 21:28:28 +00:00
|
|
|
|
|
|
|
|
2020-12-07 20:48:15 +00:00
|
|
|
def total_error(positions: numpy.ndarray, data: numpy.ndarray) -> numpy.ndarray:
|
|
|
|
"""Computes the total squared error for speaker position matrix vs data."""
|
|
|
|
return numpy.sum(numpy.square(positions - data), axis=-1)
|
2020-08-13 21:08:50 +00:00
|
|
|
|
2020-08-16 22:15:30 +00:00
|
|
|
|
2020-08-25 19:46:00 +00:00
|
|
|
@functools.lru_cache(None)
|
2020-10-15 12:08:33 +00:00
|
|
|
def frame_horizon(frame_offset: int, lookahead_steps: int):
|
2020-12-28 22:42:34 +00:00
|
|
|
"""Optimize frame_offset when more than lookahead_steps from end of frame.
|
2020-08-25 19:46:00 +00:00
|
|
|
|
2020-12-28 22:42:34 +00:00
|
|
|
Candidate opcodes for all values of frame_offset are equal, until the
|
|
|
|
end-of-frame opcode comes within our lookahead horizon.
|
2020-08-25 19:46:00 +00:00
|
|
|
"""
|
2020-12-07 20:48:15 +00:00
|
|
|
# TODO: This could be made tighter because a step is always at least 5
|
|
|
|
# cycles towards lookahead_steps.
|
2020-08-25 19:46:00 +00:00
|
|
|
if frame_offset < (2047 - lookahead_steps):
|
|
|
|
return 0
|
|
|
|
return frame_offset
|
|
|
|
|
|
|
|
|
2022-05-16 20:11:17 +00:00
|
|
|
class Speaker:
|
|
|
|
def __init__(self, sample_rate: float, freq: float, damping: float):
|
|
|
|
self.sample_rate = sample_rate
|
|
|
|
self.freq = freq
|
|
|
|
self.damping = damping
|
|
|
|
|
|
|
|
dt = numpy.float64(1 / sample_rate)
|
|
|
|
|
|
|
|
w = numpy.float64(freq * 2 * numpy.pi * dt)
|
|
|
|
|
|
|
|
d = damping * dt
|
|
|
|
e = numpy.exp(d)
|
|
|
|
c1 = 2 * e * numpy.cos(w)
|
|
|
|
|
|
|
|
c2 = e * e
|
|
|
|
t0 = (1 - 2 * e * numpy.cos(w) + e * e) / (d * d + w * w)
|
|
|
|
t = d * d + w * w - numpy.pi * numpy.pi
|
|
|
|
t1 = (1 + 2 * e * numpy.cos(w) + e * e) / numpy.sqrt(t * t + 4 * d * d *
|
|
|
|
numpy.pi * numpy.pi)
|
|
|
|
b2 = (t1 - t0) / (t1 + t0)
|
|
|
|
b1 = b2 * dt * dt * (t0 + t1) / 2
|
|
|
|
|
|
|
|
self.c1 = c1
|
|
|
|
self.c2 = c2
|
|
|
|
self.b1 = b1
|
|
|
|
self.b2 = b2
|
|
|
|
# print(dt, w, d, e, c1,c2,b1,b2)
|
|
|
|
|
|
|
|
self.scale = numpy.float64(1 / 1000) # TODO: analytic expression
|
|
|
|
|
|
|
|
def evolve(self, y1, y2, voltage1, voltage2, voltages):
|
|
|
|
output = numpy.zeros_like(voltages, dtype=numpy.float64)
|
|
|
|
x1 = numpy.full((1, voltages.shape[0]), voltage1,
|
|
|
|
dtype=numpy.float32)
|
|
|
|
x2 = numpy.full((1, voltages.shape[0]), voltage2,
|
|
|
|
dtype=numpy.float32)
|
|
|
|
for i in range(voltages.shape[1]):
|
|
|
|
# print(i)
|
|
|
|
y = self.c1 * y1 - self.c2 * y2 + self.b1 * x1 + self.b2 * x2
|
|
|
|
output[:, i] = y
|
|
|
|
|
|
|
|
y2 = y1
|
|
|
|
y1 = y
|
|
|
|
x2 = x1
|
|
|
|
x1 = voltages[:, i] # XXX does this really always lag?
|
|
|
|
|
|
|
|
# print(output)
|
|
|
|
return output
|
|
|
|
|
|
|
|
|
2020-12-28 12:29:05 +00:00
|
|
|
def audio_bytestream(data: numpy.ndarray, step: int, lookahead_steps: int,
|
2020-12-28 13:23:57 +00:00
|
|
|
sample_rate: int, is_6502: bool):
|
2020-08-16 22:15:30 +00:00
|
|
|
"""Computes optimal sequence of player opcodes to reproduce audio data."""
|
|
|
|
|
2020-08-10 20:03:12 +00:00
|
|
|
dlen = len(data)
|
2020-12-07 20:48:15 +00:00
|
|
|
# Leave enough padding at the end to look ahead from the last data value,
|
2020-12-28 22:42:34 +00:00
|
|
|
# and in case we schedule an end-of-frame opcode towards the end.
|
2020-08-25 19:46:00 +00:00
|
|
|
# TODO: avoid temporarily doubling memory footprint to concatenate
|
2020-12-07 20:48:15 +00:00
|
|
|
data = numpy.ascontiguousarray(numpy.concatenate(
|
|
|
|
[data, numpy.zeros(max(lookahead_steps, opcodes.cycle_length(
|
2020-12-28 22:42:34 +00:00
|
|
|
opcodes.Opcode.END_OF_FRAME, is_6502)), dtype=numpy.float32)]))
|
2020-08-10 20:03:12 +00:00
|
|
|
|
2020-12-07 20:48:15 +00:00
|
|
|
# Starting speaker position and applied voltage.
|
2022-05-16 20:11:17 +00:00
|
|
|
# position = 0.0
|
|
|
|
voltage1 = voltage2 = -1.0
|
2020-08-10 20:03:12 +00:00
|
|
|
|
2020-12-28 12:29:05 +00:00
|
|
|
toggles = 0
|
2022-05-16 20:11:17 +00:00
|
|
|
|
|
|
|
sp = Speaker(sample_rate, freq=3875, damping=-1210)
|
|
|
|
#
|
|
|
|
# print(sp.evolve(0, 0, 1.0, 1.0, numpy.full((1, 10000), 1.0)) * sp.scale)
|
|
|
|
# assert False
|
|
|
|
|
|
|
|
# all_partial_positions = {}
|
|
|
|
# # Precompute partial_positions so we don't skew ETA during encoding.
|
|
|
|
# for i in range(2048):
|
|
|
|
# for voltage in [-1.0, 1.0]:
|
|
|
|
# opcode_hash, _, voltages = opcodes.candidate_opcodes(
|
|
|
|
# frame_horizon(i, lookahead_steps), lookahead_steps, is_6502)
|
|
|
|
# print(i, voltages.shape[0])
|
|
|
|
# delta_powers, partial_positions = _partial_positions(
|
|
|
|
# voltage * voltages, step)
|
|
|
|
#
|
|
|
|
# # These matrices usually have more rows than columns, so store
|
|
|
|
# # then in column-major order which optimizes for this.
|
|
|
|
# delta_powers = numpy.asfortranarray(delta_powers)
|
|
|
|
# partial_positions = numpy.asfortranarray(
|
|
|
|
# partial_positions)
|
|
|
|
#
|
|
|
|
# all_partial_positions[opcode_hash, voltage] = (
|
|
|
|
# delta_powers, partial_positions)
|
|
|
|
#
|
|
|
|
# opcode_partial_positions = {}
|
|
|
|
# all_opcodes = opcodes.Opcode.__members__.values()
|
|
|
|
# for op in set(all_opcodes) - {opcodes.Opcode.EXIT}:
|
|
|
|
# voltages = opcodes.voltage_schedule(op, is_6502)
|
|
|
|
# for voltage in [-1.0, 1.0]:
|
|
|
|
# delta_powers, partial_positions = _partial_positions(
|
|
|
|
# voltage * voltages, step)
|
|
|
|
# assert delta_powers.shape == partial_positions.shape
|
|
|
|
# assert delta_powers.shape[-1] == opcodes.cycle_length(op, is_6502)
|
|
|
|
# opcode_partial_positions[op, voltage] = (
|
|
|
|
# delta_powers, partial_positions, voltage * voltages[-1])
|
|
|
|
|
|
|
|
# XXX
|
|
|
|
# Smoothing window N --> log_2 N bit resolution
|
|
|
|
# - 64
|
|
|
|
# Maintain last N voltages
|
|
|
|
# Lookahead window L
|
|
|
|
# Compute all opcodes for window L
|
|
|
|
# Compute all voltage schedules for window L
|
|
|
|
# Compute moving average over combined voltage schedule and minimize error
|
|
|
|
# XXX band pass filter first - to speaker range? no point trying to
|
|
|
|
# model frequencies that can't be produced
|
|
|
|
|
|
|
|
# old method was basically an exponential moving average, another way of
|
|
|
|
# smoothing square waveform
|
2020-12-07 20:48:15 +00:00
|
|
|
|
2020-12-07 20:58:18 +00:00
|
|
|
total_err = 0.0 # Total squared error of audio output
|
2020-12-07 20:48:15 +00:00
|
|
|
frame_offset = 0 # Position in 2048-byte TCP frame
|
2020-12-07 20:58:18 +00:00
|
|
|
i = 0 # index within input data
|
2020-12-07 20:48:15 +00:00
|
|
|
eta = ETA(total=1000, min_ms_between_updates=0)
|
|
|
|
next_tick = 0 # Value of i at which we should next update eta
|
|
|
|
# Keep track of how many opcodes we schedule
|
2020-08-16 22:15:30 +00:00
|
|
|
opcode_counts = collections.defaultdict(int)
|
2022-05-16 20:11:17 +00:00
|
|
|
|
|
|
|
y1 = y2 = 0.0 # last 2 speaker positions
|
2020-10-15 12:08:33 +00:00
|
|
|
while i < int(dlen / 1):
|
2022-05-16 20:11:17 +00:00
|
|
|
# print(i, dlen)
|
2020-12-07 20:48:15 +00:00
|
|
|
if i >= next_tick:
|
2020-08-11 22:23:33 +00:00
|
|
|
eta.print_status()
|
2020-12-07 20:48:15 +00:00
|
|
|
next_tick = int(eta.i * dlen / 1000)
|
2020-08-11 22:23:33 +00:00
|
|
|
|
2020-12-07 20:48:15 +00:00
|
|
|
# Compute all possible opcode sequences for this frame offset
|
2022-05-16 20:11:17 +00:00
|
|
|
opcode_hash, candidate_opcodes, voltages = opcodes.candidate_opcodes(
|
2020-12-28 13:23:57 +00:00
|
|
|
frame_horizon(frame_offset, lookahead_steps), lookahead_steps,
|
|
|
|
is_6502)
|
2022-05-16 20:11:17 +00:00
|
|
|
|
|
|
|
all_positions = sp.evolve(y1, y2, voltage1, voltage2, voltage1
|
|
|
|
* voltages)
|
|
|
|
# print(all_positions, all_positions.shape)
|
|
|
|
|
2020-12-07 20:48:15 +00:00
|
|
|
# Look up the precomputed partial values for these candidate opcode
|
|
|
|
# sequences.
|
2022-05-16 20:11:17 +00:00
|
|
|
# delta_powers, partial_positions = all_partial_positions[opcode_hash,
|
|
|
|
# voltage]
|
|
|
|
# # Compute matrix of new speaker positions for candidate opcode
|
|
|
|
# # sequences.
|
|
|
|
# all_positions = new_positions(position, partial_positions, delta_powers)
|
|
|
|
|
|
|
|
# opcode_idx, _ = lookahead.moving_average(
|
|
|
|
# smoothed_window, voltage * voltages, data[i:i + lookahead_steps],
|
|
|
|
# lookahead_steps)
|
2020-12-07 20:48:15 +00:00
|
|
|
|
|
|
|
assert all_positions.shape[1] == lookahead_steps
|
|
|
|
# Pick the opcode sequence that minimizes the total squared error
|
|
|
|
# relative to the data waveform. This total_error() call is where
|
|
|
|
# about 75% of CPU time is spent.
|
|
|
|
opcode_idx = numpy.argmin(
|
2022-05-16 20:11:17 +00:00
|
|
|
total_error(
|
|
|
|
all_positions * sp.scale, data[i:i + lookahead_steps])).item()
|
2020-12-07 20:48:15 +00:00
|
|
|
# Next opcode
|
2020-08-25 19:46:00 +00:00
|
|
|
opcode = candidate_opcodes[opcode_idx][0]
|
2020-12-28 13:23:57 +00:00
|
|
|
opcode_length = opcodes.cycle_length(opcode, is_6502)
|
2020-08-16 22:15:30 +00:00
|
|
|
opcode_counts[opcode] += 1
|
2020-12-28 12:29:05 +00:00
|
|
|
toggles += opcodes.TOGGLES[opcode]
|
2020-08-13 21:08:50 +00:00
|
|
|
|
2020-12-07 20:48:15 +00:00
|
|
|
# Apply this opcode to evolve the speaker position
|
2022-05-16 20:11:17 +00:00
|
|
|
opcode_voltages = (voltage1 * opcodes.voltage_schedule(
|
|
|
|
opcode, is_6502)).reshape((1, -1))
|
|
|
|
all_positions = sp.evolve(y1, y2, voltage1, voltage2, opcode_voltages)
|
|
|
|
|
|
|
|
# delta_powers, partial_positions, last_voltage = \
|
|
|
|
# opcode_partial_positions[opcode, voltage]
|
|
|
|
# all_positions = new_positions(position, partial_positions, delta_powers)
|
|
|
|
assert all_positions.shape[0] == 1
|
|
|
|
assert all_positions.shape[1] == opcode_length
|
|
|
|
|
|
|
|
voltage1 = opcode_voltages[0, -1]
|
|
|
|
voltage2 = opcode_voltages[0, -2]
|
|
|
|
y1 = all_positions[0, -1]
|
|
|
|
y2 = all_positions[0, -2]
|
|
|
|
# print(y1, y2, all_positions[0] * sp.scale)
|
2020-12-07 20:48:15 +00:00
|
|
|
total_err += total_error(
|
2022-05-16 20:11:17 +00:00
|
|
|
all_positions[0] * sp.scale, data[i:i + opcode_length]).item()
|
|
|
|
# print(all_positions[0] * sp.scale, data[i:i + opcode_length])
|
2020-12-07 20:48:15 +00:00
|
|
|
|
2022-05-16 20:11:17 +00:00
|
|
|
for v in all_positions[0]:
|
|
|
|
yield v * sp.scale
|
|
|
|
# print(v * sp.scale)
|
2020-08-13 21:08:50 +00:00
|
|
|
|
2020-12-07 20:48:15 +00:00
|
|
|
i += opcode_length
|
2020-08-13 21:08:50 +00:00
|
|
|
frame_offset = (frame_offset + 1) % 2048
|
|
|
|
|
2020-12-07 20:48:15 +00:00
|
|
|
# Make sure we have at least 2k left in stream so player will do a
|
|
|
|
# complete read.
|
2022-05-16 20:11:17 +00:00
|
|
|
# for _ in range(frame_offset % 2048, 2048):
|
|
|
|
# yield opcodes.Opcode.EXIT
|
2020-08-11 22:23:33 +00:00
|
|
|
eta.done()
|
2020-08-10 20:03:12 +00:00
|
|
|
print("Total error %f" % total_err)
|
2020-12-28 12:29:05 +00:00
|
|
|
toggles_per_sec = toggles / dlen * sample_rate
|
|
|
|
print("%d speaker toggles/sec" % toggles_per_sec)
|
2020-08-10 20:03:12 +00:00
|
|
|
|
2020-08-16 22:15:30 +00:00
|
|
|
print("Opcodes used:")
|
|
|
|
for v, k in sorted(list(opcode_counts.items()), key=lambda kv: kv[1],
|
|
|
|
reverse=True):
|
|
|
|
print("%s: %d" % (v, k))
|
|
|
|
|
2020-08-10 20:03:12 +00:00
|
|
|
|
2020-08-11 22:23:33 +00:00
|
|
|
def preprocess(
|
2020-12-28 22:42:34 +00:00
|
|
|
filename: str, target_sample_rate: int, normalize: float,
|
|
|
|
normalization_percentile: int) -> numpy.ndarray:
|
2020-08-16 22:15:30 +00:00
|
|
|
"""Upscale input audio to target sample rate and normalize signal."""
|
|
|
|
|
2020-08-11 22:23:33 +00:00
|
|
|
data, _ = librosa.load(filename, sr=target_sample_rate, mono=True)
|
|
|
|
|
2020-10-15 12:08:33 +00:00
|
|
|
max_value = numpy.percentile(data, normalization_percentile)
|
2020-08-11 22:23:33 +00:00
|
|
|
data /= max_value
|
|
|
|
data *= normalize
|
|
|
|
|
|
|
|
return data
|
|
|
|
|
2020-08-13 21:08:50 +00:00
|
|
|
|
2022-05-16 20:11:17 +00:00
|
|
|
import soundfile as sf
|
|
|
|
|
|
|
|
|
2020-12-28 12:54:44 +00:00
|
|
|
def main():
|
|
|
|
parser = argparse.ArgumentParser()
|
|
|
|
parser.add_argument("--clock", choices=['pal', 'ntsc'],
|
|
|
|
help="Whether target machine clock speed is PAL ("
|
|
|
|
"1015657Hz) or NTSC (1020484)",
|
|
|
|
required=True)
|
|
|
|
# TODO: implement 6502
|
|
|
|
parser.add_argument("--cpu", choices=['6502', '65c02'], default='65c02',
|
|
|
|
help="Target machine CPU type")
|
|
|
|
parser.add_argument("--step_size", type=int,
|
|
|
|
help="Delta encoding step size")
|
2020-12-28 22:42:34 +00:00
|
|
|
# TODO: if we're not looking ahead beyond the longest (non-end-of-frame)
|
2020-12-29 14:33:22 +00:00
|
|
|
# opcode then this will reduce quality, e.g. two opcodes may truncate to
|
|
|
|
# the same prefix, but have different results when we apply them
|
|
|
|
# fully.
|
2020-12-28 12:54:44 +00:00
|
|
|
parser.add_argument("--lookahead_cycles", type=int,
|
|
|
|
help="Number of clock cycles to look ahead in audio "
|
|
|
|
"stream.")
|
2020-12-28 22:42:34 +00:00
|
|
|
parser.add_argument("--normalization", default=1.0, type=float,
|
|
|
|
help="Overall multiplier to rescale input audio "
|
|
|
|
"values.")
|
|
|
|
parser.add_argument("--norm_percentile", default=99,
|
|
|
|
help="Normalize to specified percentile value of input "
|
|
|
|
"audio")
|
2020-12-28 12:54:44 +00:00
|
|
|
parser.add_argument("input", type=str, help="input audio file to convert")
|
|
|
|
parser.add_argument("output", type=str, help="output audio file")
|
|
|
|
args = parser.parse_args()
|
2020-08-10 20:03:12 +00:00
|
|
|
|
2020-12-24 14:43:00 +00:00
|
|
|
# Effective clock rate, including every-65 cycle "long cycle" that takes
|
|
|
|
# 16/14 as long.
|
2020-12-28 12:54:44 +00:00
|
|
|
sample_rate = 1015657 if args.clock == 'pal' else 1020484 # NTSC
|
2020-08-16 22:15:30 +00:00
|
|
|
|
2022-05-16 20:11:17 +00:00
|
|
|
# with open(args.output, "wb+") as f:[d20+
|
|
|
|
output = numpy.array(list(audio_bytestream(
|
|
|
|
preprocess(args.input, sample_rate, args.normalization,
|
|
|
|
args.norm_percentile), args.step_size,
|
|
|
|
args.lookahead_cycles, sample_rate, args.cpu == '6502')),
|
|
|
|
dtype=numpy.float32)
|
|
|
|
output_rate = 44100 # int(sample_rate / 4)
|
|
|
|
output = librosa.resample(output, orig_sr=sample_rate,
|
|
|
|
target_sr=output_rate)
|
|
|
|
with sf.SoundFile(
|
|
|
|
args.output, "w", output_rate, channels=1, format='WAV') \
|
|
|
|
as f:
|
|
|
|
f.write(output)
|
|
|
|
# f.write(bytes([opcode.value]))
|
2020-08-10 20:03:12 +00:00
|
|
|
|
|
|
|
|
|
|
|
if __name__ == "__main__":
|
2020-12-28 12:54:44 +00:00
|
|
|
main()
|