Inconsistent mask shape in forward_tts _forward_aligner and AlignmentNetwork docstrings

[Prajwalbhandari869]

In the docstrings of forward_tts _forward_aligner, it is said the shape of parameter x_mask should be as - x_mask: :math:[B, 1, T_en] and the same shaped x_mask is passed to aligner without squeezing.

coqui-ai-TTS/TTS/tts/models/forward_tts.py

Line 507 in b20533e

- x_mask: :math:`[B, 1, T_en]`

But forward of AlignmentNetwork is expecting x_mask as - mask: :math:[B, T_de].

coqui-ai-TTS/TTS/tts/layers/generic/aligner.py

Line 74 in b20533e

- mask: :math:`[B, T_de]`

So, what is the actual size expected at AlignmentNetwork forward?

• Should it be actual passed - x_mask: :math:[B, 1, T_en]?
• Or Should it be expecting - mask: :math:[B, T_de]?
• Or is the docstring inconsistent?

[eginhard]

I wouldn't rely on docstrings in Coqui to be consistent, but PRs with fixes are always welcome! Inserting print("x_mask", x_mask.shape) after

coqui-ai-TTS/TTS/tts/models/forward_tts.py

Line 647 in b20533e

o_en, x_mask, g, _ = self._forward_encoder(x, x_mask, g)

prints x_mask torch.Size([1, 1, 10]) when running tts --model_name "tts_models/en/ljspeech/fast_pitch" --text "hello world" (that's the inference function but it should be the same in forward).

Also, in the future please share code as text or link directly to the correct location (I've edited your question accordingly). Images make it difficult to locate the code and aren't accessible for people with visual impairments.

Old repo crosslink: coqui-ai#4158

[Prajwalbhandari869]

Later, I checked the code with a tensor, and it was okay. Thank you for your reply.