mstar.model.omnilingual_asr.config

mstar.model.omnilingual_asr.config#

Omnilingual ASR CTC: architecture per size, and where each checkpoint lives.

Every size shares one code path; only the encoder’s width and depth and the vocabulary differ (fairseq2 wav2vec2_asr archs 300m/1b/3b/7b and their _v2 twins, in omnilingual_asr/models/wav2vec2_asr/config.py).

Weights come in two releases, both Apache-2.0:

  • v2 (default): Meta’s December 2025 checkpoints, lower CER per Meta, with a “written” vocabulary that adds capitals and punctuation. Published only on Meta’s CDN (dl.fbaipublicfiles.com/mms/), with no checksums. The sha256 values below were computed by us from the downloads (October 2026).

  • v1: the November 2025 checkpoints, on the Hugging Face Hub as facebook/omniASR-CTC-*, pinned to a revision; the sha256 values are the Hub’s LFS hashes.

Every file fetched by name is checked against its hash before it is loaded, whether downloaded now or found in the cache, so a replaced file fails loudly.

Functions

ctc_config(size, version)

parse_dtype(dtype)

Classes

Artifact(url, sha256, filename)

HubCheckpoint(repo_id, revision, filename, ...)

OmniASRCTCConfig(model_dim, num_layers, ...)

class mstar.model.omnilingual_asr.config.Artifact(url: 'str', sha256: 'str', filename: 'str')[source]#

Bases: object

Parameters:
filename: str#
sha256: str#
url: str#
class mstar.model.omnilingual_asr.config.HubCheckpoint(repo_id: 'str', revision: 'str', filename: 'str', sha256: 'str')[source]#

Bases: object

Parameters:
filename: str#
repo_id: str#
revision: str#
sha256: str#
class mstar.model.omnilingual_asr.config.OmniASRCTCConfig(model_dim: 'int', num_layers: 'int', ffn_inner_dim: 'int', vocab_size: 'int', num_heads: 'int' = 16, feature_dim: 'int' = 512, feature_extractor_layer_descs: 'list[tuple[int, int, int]]'=<factory>, pos_conv_kernel_size: 'int' = 128, num_pos_conv_groups: 'int' = 16, layer_norm_eps: 'float' = 1e-05)[source]#

Bases: object

Parameters:

  • model_dim (int)

  • num_layers (int)

  • ffn_inner_dim (int)

  • vocab_size (int)

  • num_heads (int)

  • feature_dim (int)

  • feature_extractor_layer_descs (list[tuple[int, int, int]])

  • pos_conv_kernel_size (int)

  • num_pos_conv_groups (int)

  • layer_norm_eps (float)

feature_dim: int = 512#
feature_extractor_layer_descs: list[tuple[int, int, int]]#
ffn_inner_dim: int#
property hop_samples: int#

50 frames/s).

Type:

Samples per encoder frame (320

layer_norm_eps: float = 1e-05#
model_dim: int#
num_heads: int = 16#
num_layers: int#
num_pos_conv_groups: int = 16#
pos_conv_kernel_size: int = 128#
vocab_size: int#
mstar.model.omnilingual_asr.config.ctc_config(size, version)[source]#
Parameters:
Return type:

OmniASRCTCConfig

mstar.model.omnilingual_asr.config.parse_dtype(dtype)[source]#
Parameters:

dtype (str | dtype)

Return type:

dtype