[WIP] Add DSP ops

examples/tutorials/fm_synthesis_tutorial.py

@@ -0,0 +1,111 @@
+"""
+FM Synthesis
+============
+
+**Author**: `Moto Hira <moto@meta.com>`__
+
+.. warning::
+   This tutorial requires prototype DSP features, which are
+   available in nightly builds.
+
+   Please refer to https://pytorch.org/get-started/locally
+   for instructions for installing a nightly build.
+
+"""
+
+import torch
+import torchaudio
+
+print(torch.__version__)
+print(torchaudio.__version__)
+
+######################################################################
+#
+from torchaudio.prototype.functional import oscillator_bank
+
+import matplotlib.pyplot as plt
+from IPython.display import Audio
+
+######################################################################
+#
+
+SAMPLE_RATE = 16000
+duration = 1.2
+NUM_FRAMES = int(SAMPLE_RATE * duration)
+
+F0 = 220.
+F1 = 440.
+
+
+def fm_synth(beta, f1=F1, f0=F0):
+    print(f"Modulator Frequency: {f1} [Hz]")
+    print(f"Carrior Frequency: {f0} [Hz]")
+    print(f"Beta: {beta}")
+
+    amp = torch.ones((NUM_FRAMES, 1), dtype=torch.float64)
+
+    freq = torch.full((NUM_FRAMES, 1), f1, dtype=torch.float64)
+    mod = oscillator_bank(freq, amp, sample_rate=SAMPLE_RATE, reduction="none")
+
+    freq = f0 + beta * f0 * mod
+    car = oscillator_bank(freq, amp, sample_rate=SAMPLE_RATE)
+    return mod.sum(-1), car
+
+
+######################################################################
+#
+def plot(mod, car, sample_rate=SAMPLE_RATE):
+    fig, axes = plt.subplots(4, 1)
+    axes[0].specgram(mod, Fs=sample_rate)
+    spectrum, freqs, _, _ = axes[1].specgram(car, Fs=sample_rate)
+
+    num_samples = int(SAMPLE_RATE * 0.01)
+    t = torch.linspace(0, num_samples / sample_rate, num_samples)
+    axes[2].plot(t, car[..., :num_samples])
+
+    half = freqs.shape[0] // 2
+    axes[3].plot(freqs[:half], spectrum[:half, spectrum.shape[1]//2], marker='+')
+
+
+######################################################################
+#
+mod, car = fm_synth(beta=0.1)
+plot(mod, car)
+Audio(car, rate=SAMPLE_RATE)
+######################################################################
+#
+mod, car = fm_synth(beta=1)
+plot(mod, car)
+Audio(car, rate=SAMPLE_RATE)
+######################################################################
+#
+mod, car = fm_synth(beta=10)
+plot(mod, car)
+Audio(car, rate=SAMPLE_RATE)
+
+
+######################################################################
+#
+mod, car = fm_synth(beta=1, f1=0.01*F0)
+plot(mod, car)
+Audio(car, rate=SAMPLE_RATE)
+######################################################################
+#
+mod, car = fm_synth(beta=1, f1=0.1*F0)
+plot(mod, car)
+Audio(car, rate=SAMPLE_RATE)
+######################################################################
+#
+mod, car = fm_synth(beta=10, f1=0.1*F0)
+plot(mod, car)
+Audio(car, rate=SAMPLE_RATE)
+######################################################################
+#
+mod, car = fm_synth(beta=1, f1=1*F0)
+plot(mod, car)
+Audio(car, rate=SAMPLE_RATE)
+######################################################################
+#
+mod, car = fm_synth(beta=1, f1=10*F0)
+plot(mod, car)
+Audio(car, rate=SAMPLE_RATE)

repro_dsp.py

@@ -0,0 +1,54 @@
+import torch
+from torchaudio.prototype.functional import oscillator_bank, adsr_envelope
+
+import matplotlib.pyplot as plt
+
+
+def osci():
+    sample_rate = 8000
+    num_frames = 8000 * 3
+
+    freq = torch.ones((num_frames, 1)) * torch.tensor([[2000]])  # , 3000]])
+    amps = torch.ones_like(freq)
+    waveform32 = oscillator_bank(freq, amps, sample_rate)
+
+    freq = freq.to(torch.float64)
+    amps = freq.to(torch.float64)
+    waveform64 = oscillator_bank(freq, amps, sample_rate)
+
+    fig, axes = plt.subplots(2, 1, sharex=True, sharey=True)
+    fig.suptitle("Precision and waveform generated by oscillator_bank")
+    _, _, _, cax = axes[0].specgram(waveform32, Fs=sample_rate)
+    axes[0].set(title="float32", ylabel="Frequency [Hz]")
+    fig.colorbar(cax)
+
+    _, _, _, cax = axes[1].specgram(waveform64, Fs=sample_rate)
+    axes[1].set(title="float64", xlabel="time [s]", ylabel="Frequency [Hz]")
+    plt.colorbar(cax)
+    plt.tight_layout()
+    plt.subplots_adjust(left=0.2, top=0.85, right=0.9,bottom=0.15)
+    fig.savefig("oscillator_precision.png", dpi=200, transparent=True)
+
+
+def adsr():
+    num_frames = 8000
+    configs = [
+        {"attack": 0.2, "hold": 0.2, "decay": 0.2, "sustain": 0.5, "release": 0.2},
+        {"attack": 0.02, "decay": 0.98, "sustain": 0, "release": 0},
+        {"attack": 0.01, "hold": 0.3, "decay": 0.05, "sustain": 0.01, "release": 0},
+        {"attack": 0.98, "decay": 0, "sustain": 1, "release": 0.02},
+    ]
+    waveforms = [adsr_envelope(**config, num_frames=num_frames) for config in configs]
+    t = torch.linspace(0, 1.0, num_frames)
+
+    fig, axes = plt.subplots(len(configs), 1, sharex=True, sharey=True)
+    for ax, config, waveform in zip(axes, configs, waveforms):
+        ax.plot(t, waveform)
+        ax.grid(True)
+        ax.set(title=', '.join(f'{k}: {v}' for k, v in config.items()))
+    fig.tight_layout()
+    fig.savefig("adsr.png", dpi=200)
+
+# osci()
+adsr()
+plt.show()