mstar.model.kokoro.config

Kokoro-82M configuration, built from the checkpoint’s config.json.

The upstream kokoro package hard-codes a handful of architecture constants outside its config (the 24 kHz sample rate and the harmonic source parameters of the iSTFTNet generator, the 1024-wide decoder trunk, the 64-channel ASR residual). They are fields here so nothing else in the package carries a magic number.

Classes

KokoroBertConfig([hidden_size, ...])	PL-BERT: an ALBERT encoder over phoneme tokens (plbert in config.json).
KokoroISTFTNetConfig([...])	The iSTFTNet generator (istftnet in config.json).
KokoroModelConfig([n_token, hidden_dim, ...])	Everything the M* Kokoro integration reads.

class mstar.model.kokoro.config.KokoroBertConfig(hidden_size=768, num_attention_heads=12, intermediate_size=2048, max_position_embeddings=512, num_hidden_layers=12, dropout=0.1, embedding_size=128, layer_norm_eps=1e-12)

Bases: object

PL-BERT: an ALBERT encoder over phoneme tokens (plbert in config.json).

embedding_size and layer_norm_eps are ALBERT defaults the checkpoint relies on without stating them.

Parameters:

• hidden_size (int)

• num_attention_heads (int)

• intermediate_size (int)

• max_position_embeddings (int)

• num_hidden_layers (int)

• dropout (float)

• embedding_size (int)

• layer_norm_eps (float)

dropout: float = 0.1

embedding_size: int = 128

property head_dim: int

hidden_size: int = 768

intermediate_size: int = 2048

layer_norm_eps: float = 1e-12

max_position_embeddings: int = 512

num_attention_heads: int = 12

num_hidden_layers: int = 12

class mstar.model.kokoro.config.KokoroISTFTNetConfig(upsample_kernel_sizes=<factory>, upsample_rates=<factory>, gen_istft_hop_size=5, gen_istft_n_fft=20, resblock_dilation_sizes=<factory>, resblock_kernel_sizes=<factory>, upsample_initial_channel=512)

Bases: object

The iSTFTNet generator (istftnet in config.json).

Parameters:

• upsample_kernel_sizes (list[int])

• upsample_rates (list[int])

• gen_istft_hop_size (int)

• gen_istft_n_fft (int)

• resblock_dilation_sizes (list[list[int]])

• resblock_kernel_sizes (list[int])

• upsample_initial_channel (int)

gen_istft_hop_size: int = 5

gen_istft_n_fft: int = 20

resblock_dilation_sizes: list[list[int]]

resblock_kernel_sizes: list[int]

property upsample_factor: int

Waveform samples per generator input frame (60 * 5 = 300).

upsample_initial_channel: int = 512

upsample_kernel_sizes: list[int]

upsample_rates: list[int]

class mstar.model.kokoro.config.KokoroModelConfig(n_token=178, hidden_dim=512, style_dim=128, n_layer=3, max_dur=50, dropout=0.2, n_mels=80, text_encoder_kernel_size=5, dim_in=64, max_conv_dim=512, multispeaker=True, vocab=<factory>, plbert=<factory>, istftnet=<factory>, sample_rate=24000, decoder_hidden=1024, asr_res_dim=64, source_harmonics=8, source_sine_amp=0.1, source_noise_std=0.003, source_voiced_threshold=10.0, weights_file='kokoro-v1_0.pth', voices_dir='voices', default_voice='af_heart', default_speed=1.0, min_speed=0.25, max_speed=4.0, first_chunk_target_phonemes=80, chunk_target_phonemes=200, max_chunks=512, text_buckets=<factory>, frame_buckets=<factory>, capture_batch_sizes=<factory>, max_batch_frames=16384, frame_grouping='bucket', decoder_dtype='float32', compile_decoder=True)

Bases: object

Everything the M* Kokoro integration reads.

The top block mirrors config.json; the serving block has M*-side defaults that a deployment can change through model kwargs.

Parameters:

• n_token (int)

• hidden_dim (int)

• style_dim (int)

• n_layer (int)

• max_dur (int)

• dropout (float)

• n_mels (int)

• text_encoder_kernel_size (int)

• dim_in (int)

• max_conv_dim (int)

• multispeaker (bool)

• vocab (dict[str, int])

• plbert (KokoroBertConfig)

• istftnet (KokoroISTFTNetConfig)

• sample_rate (int)

• decoder_hidden (int)

• asr_res_dim (int)

• source_harmonics (int)

• source_sine_amp (float)

• source_noise_std (float)

• source_voiced_threshold (float)

• weights_file (str)

• voices_dir (str)

• default_voice (str)

• default_speed (float)

• min_speed (float)

• max_speed (float)

• first_chunk_target_phonemes (int)

• chunk_target_phonemes (int)

• max_chunks (int)

• text_buckets (list[int])

• frame_buckets (list[int])

• capture_batch_sizes (list[int])

• max_batch_frames (int)

• frame_grouping (str)

• decoder_dtype (str)

• compile_decoder (bool)

asr_res_dim: int = 64

capture_batch_sizes: list[int]

chunk_target_phonemes: int = 200

compile_decoder: bool = True

property context_length: int

decoder_dtype: str = 'float32'

decoder_hidden: int = 1024

default_speed: float = 1.0

default_voice: str = 'af_heart'

dim_in: int = 64

dropout: float = 0.2

first_chunk_target_phonemes: int = 80

frame_buckets: list[int]

frame_grouping: str = 'bucket'

classmethod from_dict(data)

Parameters:

data (dict[str, Any])

Return type:

KokoroModelConfig

classmethod from_pretrained(local_dir)

Parameters:

local_dir (str | Path)

Return type:

KokoroModelConfig

hidden_dim: int = 512

istftnet: KokoroISTFTNetConfig

max_batch_frames: int = 16384

max_chunks: int = 512

max_conv_dim: int = 512

max_dur: int = 50

property max_phonemes: int

Phonemes per chunk, leaving room for the <bos>/<eos> pair.

max_speed: float = 4.0

min_speed: float = 0.25

multispeaker: bool = True

n_layer: int = 3

n_mels: int = 80

n_token: int = 178

plbert: KokoroBertConfig

sample_rate: int = 24000

property samples_per_frame: int

the prosody predictor doubles the frame rate, the generator upsamples by 300.

Type:

Waveform samples per predicted duration unit

source_harmonics: int = 8

source_noise_std: float = 0.003

source_sine_amp: float = 0.1

source_voiced_threshold: float = 10.0

style_dim: int = 128

property style_pack_rows: int

A voice pack holds one style vector per phoneme count.

property style_vector_dim: int

A pack row is [decoder style, predictor style].

text_buckets: list[int]

text_encoder_kernel_size: int = 5

vocab: dict[str, int]

voices_dir: str = 'voices'

weights_file: str = 'kokoro-v1_0.pth'