---

Waveform Synthesis of Held-out LibriSpeech Samples

Original audio
Griffin-Lim (3 iterations)
Griffin-Lim (50 iterations)
Griffin-Lim (150 iterations)
SPSI
SPSI + Griffin-Lim (3 iterations)
SPSI + Griffin-Lim (50 iterations)
MCNN (baseline)
MCNN (2 heads)
MCNN (filter width of 9)
MCNN (losses (1) and (2))
MCNN (loss (1) only)

Original audio
Griffin-Lim (3 iterations)
Griffin-Lim (50 iterations)
Griffin-Lim (150 iterations)
SPSI
SPSI + Griffin-Lim (3 iterations)
SPSI + Griffin-Lim (50 iterations)
MCNN (baseline)
MCNN (2 heads)
MCNN (filter width of 9)
MCNN (losses (1) and (2))
MCNN (loss (1) only)

Original audio
Griffin-Lim (3 iterations)
Griffin-Lim (50 iterations)
Griffin-Lim (150 iterations)
SPSI
SPSI + Griffin-Lim (3 iterations)
SPSI + Griffin-Lim (50 iterations)
MCNN (baseline)
MCNN (2 heads)
MCNN (filter width of 9)
MCNN (losses (1) and (2))
MCNN (loss (1) only)

---

Waveform Synthesis of Samples from an Unseen Speaker

Original audio
Griffin-Lim (150 iterations)
SPSI + Griffin-Lim (50 iterations)
MCNN (baseline)
MCNN (trained on the unseen speaker dataset)

Original audio
Griffin-Lim (150 iterations)
SPSI + Griffin-Lim (50 iterations)
MCNN (baseline)
MCNN (trained on the unseen speaker dataset)

Original audio
Griffin-Lim (150 iterations)
SPSI + Griffin-Lim (50 iterations)
MCNN (baseline)
MCNN (trained on the unseen speaker dataset)

---

Contributions of Multiple Heads

Original audio
MCNN
Head 1 output
Head 2 output
Head 3 output
Head 4 output
Head 5 output
Head 6 output
Head 7 output
Head 8 output