geolid_combined_shared_trainable

ESPnet2 Spoken Language Identification (LID) model

espnet/geolid_combined_shared_trainable

Paper

This geolocation-aware language identification (LID) model is developed using the ESPnet toolkit. It integrates the powerful pretrained MMS-1B as the encoder and employs ECAPA-TDNN as the embedding extractor to achieve robust spoken language identification.

The main innovations of this model are:

1. Incorporating geolocation prediction as an auxiliary task during training.
2. Conditioning the intermediate representations of the self-supervised learning (SSL) encoder on intermediate-layer information. This geolocation-aware strategy greatly improves robustness, especially for dialects and accented variations.

For further details on the geolocation-aware LID methodology, please refer to our paper: Geolocation-Aware Robust Spoken Language Identification (arXiv).

Usage Guide: How to use in ESPnet2

Prerequisites

First, ensure you have ESPnet installed. If not, follow the ESPnet installation instructions.

Quick Start

Run the following commands to set up and use the pre-trained model:

cd espnet

pip install -e .
cd egs2/geolid/lid1

# Download the exp_combined to egs2/geolid/lid1
# Make sure hf CLI is installed: pip install -U "huggingface_hub[cli]"
hf download espnet/geolid_combined_shared_trainable --local-dir . --exclude "README.md" "meta.yaml" ".gitattributes"

./run_combined.sh --skip_data_prep false --skip_train true

This will download the pre-trained model and run inference.

Train and Evaluation Datasets

The training utilized a combined dataset, merging five domain-specific corpora, resulting in 9,865 hours of speech data covering 157 languages.

Results

Accuracy (%) on In-domain and Out-of-domain Test Sets

<style> .hf-model-cell { max-width: 120px; overflow-x: auto; white-space: nowrap; scrollbar-width: thin; scrollbar-color: #888 #f1f1f1; } .config-cell { max-width: 100px; overflow-x: auto; white-space: nowrap; scrollbar-width: thin; scrollbar-color: #888 #f1f1f1; } .hf-model-cell::-webkit-scrollbar, .config-cell::-webkit-scrollbar { height: 6px; } .hf-model-cell::-webkit-scrollbar-track, .config-cell::-webkit-scrollbar-track { background: #f1f1f1; border-radius: 3px; } .hf-model-cell::-webkit-scrollbar-thumb, .config-cell::-webkit-scrollbar-thumb { background: #888; border-radius: 3px; } .hf-model-cell::-webkit-scrollbar-thumb:hover, .config-cell::-webkit-scrollbar-thumb:hover { background: #555; } </style> <div style="overflow-x: auto;"> </div>

For more detailed inference results, please refer to the exp_combined/lid_mms_ecapa_upcon_32_44_it0.4_shared_trainable_raw/inference directory in this repository.

Note (2025-08-18):
The corresponding GitHub recipe egs2/geolid/lid1 has not yet been merged into the ESPnet master branch.
See TODO: add PR link for the latest updates.

LID config

<details><summary>expand</summary>

config: conf/combined/mms_ecapa_upcon_32_44_it0.4_shared_trainable_dev.yaml
print_config: false
log_level: INFO
drop_last_iter: false
dry_run: false
iterator_type: category
valid_iterator_type: category
output_dir: exp_combined/lid_mms_ecapa_upcon_32_44_it0.4_shared_trainable_dev_raw
ngpu: 1
seed: 3702
num_workers: 8
num_att_plot: 0
dist_backend: nccl
dist_init_method: env://
dist_world_size: null
dist_rank: null
local_rank: 0
dist_master_addr: null
dist_master_port: null
dist_launcher: null
multiprocessing_distributed: false
unused_parameters: true
sharded_ddp: false
use_deepspeed: false
deepspeed_config: null
gradient_as_bucket_view: true
ddp_comm_hook: null
cudnn_enabled: true
cudnn_benchmark: true
cudnn_deterministic: false
use_tf32: false
collect_stats: false
write_collected_feats: false
max_epoch: 33
patience: null
val_scheduler_criterion:
- valid
- loss
early_stopping_criterion:
- valid
- loss
- min
best_model_criterion:
-   - valid
    - accuracy
    - max
keep_nbest_models: 2
nbest_averaging_interval: 0
grad_clip: 9999
grad_clip_type: 2.0
grad_noise: false
accum_grad: 4
no_forward_run: false
resume: true
train_dtype: float32
use_amp: true
log_interval: 100
use_matplotlib: true
use_tensorboard: true
create_graph_in_tensorboard: false
use_wandb: false
wandb_project: null
wandb_id: null
wandb_entity: null
wandb_name: null
wandb_model_log_interval: -1
detect_anomaly: false
use_adapter: false
adapter: lora
save_strategy: all
adapter_conf: {}
pretrain_path: null
init_param: []
ignore_init_mismatch: false
freeze_param: []
num_iters_per_epoch: 2000
batch_size: 20
valid_batch_size: null
batch_bins: 1440000
valid_batch_bins: null
category_sample_size: 10
upsampling_factor: 0.5
category_upsampling_factor: 0.5
dataset_upsampling_factor: 0.3
dataset_scaling_factor: 1.2
max_batch_size: 6
min_batch_size: 1
train_shape_file:
- exp_combined/lid_stats_16k/train/speech_shape
valid_shape_file:
- exp_combined/lid_stats_16k/valid/speech_shape
batch_type: catpow_balance_dataset
language_upsampling_factor: 0.5
valid_batch_type: null
fold_length:
- 120000
sort_in_batch: descending
shuffle_within_batch: false
sort_batch: descending
multiple_iterator: false
chunk_length: 500
chunk_shift_ratio: 0.5
num_cache_chunks: 1024
chunk_excluded_key_prefixes: []
chunk_default_fs: null
chunk_max_abs_length: null
chunk_discard_short_samples: true
train_data_path_and_name_and_type:
-   - dump/raw/train_all_no_filter_lang/wav.scp
    - speech
    - sound
-   - dump/raw/train_all_no_filter_lang/utt2lang
    - lid_labels
    - text
valid_data_path_and_name_and_type:
-   - dump/raw/dev_ml_superb2_lang/wav.scp
    - speech
    - sound
-   - dump/raw/dev_ml_superb2_lang/utt2lang
    - lid_labels
    - text
multi_task_dataset: false
allow_variable_data_keys: false
max_cache_size: 0.0
max_cache_fd: 32
allow_multi_rates: false
valid_max_cache_size: null
exclude_weight_decay: false
exclude_weight_decay_conf: {}
optim: adam
optim_conf:
    lr: 1.0e-05
    betas:
    - 0.9
    - 0.98
scheduler: tristagelr
scheduler_conf:
    max_steps: 12500
    warmup_ratio: 0.1
    hold_ratio: 0.4
    decay_ratio: 0.5
    init_lr_scale: 0.6
    final_lr_scale: 0.1
init: null
use_preprocessor: true
input_size: null
target_duration: 3.0
lang2utt: dump/raw/train_all_no_filter_lang/lang2utt
lang_num: 157
sample_rate: 16000
num_eval: 10
rir_scp: ''
model: upstream_condition
model_conf:
    lang2vec_conditioning_layers:
    - 32
    - 36
    - 40
    - 44
    apply_intermediate_lang2vec_loss: true
    apply_intermediate_lang2vec_condition: true
    inter_lang2vec_loss_weight: 0.4
    cutoff_gradient_from_backbone: false
    cutoff_gradient_before_condproj: true
    shared_conditioning_proj: true
frontend: s3prl_condition
frontend_conf:
    frontend_conf:
        upstream: hf_wav2vec2_condition
        path_or_url: facebook/mms-1b
    download_dir: ./hub
    multilayer_feature: true
specaug: null
specaug_conf: {}
normalize: utterance_mvn
normalize_conf:
    norm_vars: false
encoder: ecapa_tdnn
encoder_conf:
    model_scale: 8
    ndim: 512
    output_size: 1536
pooling: chn_attn_stat
pooling_conf: {}
projector: rawnet3
projector_conf:
    output_size: 192
encoder_condition: identity
encoder_condition_conf: {}
pooling_condition: chn_attn_stat
pooling_condition_conf: {}
projector_condition: rawnet3
projector_condition_conf: {}
preprocessor: lid
preprocessor_conf:
    fix_duration: false
    sample_rate: 16000
    noise_apply_prob: 0.0
    noise_info:
    -   - 1.0
        - dump/raw/musan_speech.scp
        -   - 4
            - 7
        -   - 13
            - 20
    -   - 1.0
        - dump/raw/musan_noise.scp
        -   - 1
            - 1
        -   - 0
            - 15
    -   - 1.0
        - dump/raw/musan_music.scp
        -   - 1
            - 1
        -   - 5
            - 15
    rir_apply_prob: 0.0
    rir_scp: dump/raw/rirs.scp
    use_lang2vec: true
    lang2vec_type: geo
loss: aamsoftmax_sc_topk_lang2vec
loss_conf:
    margin: 0.5
    scale: 30
    K: 3
    mp: 0.06
    k_top: 5
    lang2vec_dim: 299
    lang2vec_type: geo
    lang2vec_weight: 0.2
required:
- output_dir
version: '202506'
distributed: false

</details>

Citation

@inproceedings{wang2025geolid,
  author={Qingzheng Wang, Hye-jin Shim, Jiancheng Sun, and Shinji Watanabe},
  title={Geolocation-Aware Robust Spoken Language Identification},
  year={2025},
  booktitle={Procedings of ASRU},
}

@inproceedings{watanabe2018espnet,
  author={Shinji Watanabe and Takaaki Hori and Shigeki Karita and Tomoki Hayashi and Jiro Nishitoba and Yuya Unno and Nelson Yalta and Jahn Heymann and Matthew Wiesner and Nanxin Chen and Adithya Renduchintala and Tsubasa Ochiai},
  title={{ESPnet}: End-to-End Speech Processing Toolkit},
  year={2018},
  booktitle={Proceedings of Interspeech},
  pages={2207--2211},
  doi={10.21437/Interspeech.2018-1456},
  url={http://dx.doi.org/10.21437/Interspeech.2018-1456}
}