mms1b-lid-transformer-pit-fleurs-csfleurs

ESPnet FLEURS-CS ASR, explicitly selected epoch (15epoch.pth). This is not a claim of best-checkpoint performance.

License: CC-BY-NC-4.0 (including the MMS-1B backbone where used). No evaluation metrics are asserted. See provenance.json for hashes.

No optimizer state, audio or real utterance IDs. Use matching ESPnet CS-LID code (canonical PIT keys); MMS inference still requires its backbone/cache.

The config retains training settings, including optimizer/scheduler hyperparameters, but no optimizer/scheduler state is uploaded. Historical training data, initialization and output paths are provenance, not bundled assets; relocate them before training on another server. Language inventory and ASR token/symbol references are made bundle-local or inlined.

LID: resolve lang2utt relative to the bundle for upstream preprocessing. Its placeholders are not training references. Optional utt2langs contains explicitly supplied frozen training language sets with synthetic IDs: valid for seen/unseen membership, never for utterance counts. Without that file, supply actual frozen training references separately.

Architecture

Selected settings from the validated portable config:

frontend: s3prl
frontend_conf:
  fs: 16k
  frontend_conf:
    upstream: hf_wav2vec2_custom
    path_or_url: facebook/mms-1b
  download_dir: ./hub
  multilayer_feature: true
encoder: transformer
encoder_conf:
  output_size: 256
  attention_heads: 4
  linear_units: 1024
  num_blocks: 24
  dropout_rate: 0.1
  positional_dropout_rate: 0.1
  attention_dropout_rate: 0.0
  input_layer: linear
  normalize_before: true
model_conf:
  extract_feats_in_collect_stats: false
  ctc_weight: 0.0
  lsm_weight: 0.1
  length_normalized_loss: false
  pit_loss: true
  pit_loss_reduction: min
freeze_param: []
decoder: transformer
decoder_conf:
  attention_heads: 4
  linear_units: 1024
  num_blocks: 4
  dropout_rate: 0.1
  positional_dropout_rate: 0.1
  self_attention_dropout_rate: 0.0
  src_attention_dropout_rate: 0.0