mstar.model.kokoro.components

class mstar.model.kokoro.components.KokoroTTS(config)

Bases: Module

Parameters:

config (KokoroModelConfig)

align(d, t_en, pred_dur, num_frames)

Expand phoneme features to num_frames frames.

Returns the frame-aligned prosody states [B, F, hidden + style], the aligned text features [B, hidden, F] and each row’s valid frame count.

Parameters:

• d (Tensor)

• t_en (Tensor)

• pred_dur (Tensor)

• num_frames (int)

Return type:

tuple[Tensor, Tensor, Tensor]

static alignment(pred_dur, num_frames)

Index of the phoneme that owns each frame, [B, num_frames].

Equal to the reference’s repeat_interleave + one-hot matmul on the valid frames; frames past a row’s total point at its last phoneme.

Parameters:

• pred_dur (Tensor)

• num_frames (int)

Return type:

Tensor

decode_frames(en, asr, frame_lengths, style)

Frame-aligned features -> waveform [B, F * samples_per_frame].

Parameters:

• en (Tensor)

• asr (Tensor)

• frame_lengths (Tensor)

• style (Tensor)

Return type:

Tensor

encode_text(input_ids, lengths, style, speed)

[B, T] ids -> prosody states [B, T, hidden + style], text features [B, hidden, T] and integer durations [B, T].

Parameters:

• input_ids (Tensor)

• lengths (Tensor)

• style (Tensor)

• speed (Tensor)

Return type:

tuple[Tensor, Tensor, Tensor]

forward(input_ids, lengths, style, speed)

Eager end-to-end synthesis: (waveform, frame lengths, durations).

Parameters:

• input_ids (Tensor)

• lengths (Tensor)

• style (Tensor)

• speed (Tensor)

Return type:

tuple[Tensor, Tensor, Tensor]

synthesize_frames(d, t_en, pred_dur, style, num_frames)

align then decode_frames: the waveform and each row’s frame count.

Parameters:

• d (Tensor)

• t_en (Tensor)

• pred_dur (Tensor)

• style (Tensor)

• num_frames (int)

Return type:

tuple[Tensor, Tensor]

Modules

adain	Style-conditioned normalization blocks shared by the prosody predictor and the decoder (StyleTTS2's AdaIN1d / AdainResBlk1d).
albert	PL-BERT: the ALBERT phoneme encoder that conditions Kokoro's prosody.
decoder	iSTFTNet decoder: prosody-conditioned convolutional trunk, harmonic-plus-noise excitation, two upsampling stages with Snake residual blocks, and an inverse STFT head (StyleTTS2's Modules/istftnet.py as adapted by Kokoro).
lstm	Bidirectional LSTM over padded batches without packing.
masking	Length masks for padded batches.
prosody	Text encoder and prosody predictor (durations, F0 and energy) of Kokoro.
stft	Short-time Fourier analysis and synthesis over padded batches.
synthesizer	The whole Kokoro network: phonemes + style -> waveform, batched and masked.