---
tags:
- espnet
- audio
- language-identification
language: 
- abk
- afr
- amh
- ara
- asm
- aze
- bak
- bel
- ben
- bod
- bos
- bre
- bul
- cat
- ceb
- ces
- cmn
- cym
- dan
- deu
- ell
- eng
- epo
- est
- eus
- fao
- fas
- fin
- fra
- glg
- glv
- grn
- guj
- hat
- hau
- haw
- heb
- hin
- hrv
- hun
- hye
- ina
- ind
- isl
- ita
- jav
- jpn
- kan
- kat
- kaz
- khm
- kor
- lao
- lat
- lav
- lin
- lit
- ltz
- mal
- mar
- mkd
- mlg
- mlt
- mon
- mri
- msa
- mya
- nep
- nld
- nno
- nor
- oci
- pan
- pol
- por
- pus
- ron
- rus
- san
- sco
- sin
- slk
- slv
- sna
- snd
- som
- spa
- sqi
- srp
- sun
- swa
- swe
- tam
- tat
- tel
- tgk
- tgl
- tha
- tuk
- tur
- ukr
- urd
- uzb
- vie
- war
- yid
- yor
datasets:
- VoxLingua107
license: cc-by-4.0
---

## ESPnet2 Spoken Language Identification (LID) model

### `espnet/lid_voxlingua107_mms_ecapa`

This language identification model was trained using the ESPnet recipe from [ESPnet](https://github.com/espnet/espnet/) toolkit. It leverages the pretrained [MMS-1B](https://huggingface.co/facebook/mms-1b) as the encoder and [ECAPA-TDNN](https://arxiv.org/pdf/2005.07143) as the embedding extractor for robust spoken language identification.

The model is trained on the [VoxLingua107](https://cs.taltech.ee/staff/tanel.alumae/data/voxlingua107/) dataset, which comprises over 6,600 hours of speech spanning 107 languages. Speech segments are sourced from YouTube videos and annotated using metadata.

This repository provides comprehensive training logs, detailed inference results, and model checkpoints for reproducibility and further research.

### Usage Guide: How to use in ESPnet2

#### Prerequisites
First, ensure you have ESPnet installed. If not, follow the [ESPnet installation instructions](https://espnet.github.io/espnet/installation.html).

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

```bash
cd espnet

pip install -e .
cd egs2/voxlingua107/lid1

# Download the exp_voxlingua107_raw to egs2/voxlingua107/lid1
hf download espnet/lid_voxlingua107_mms_ecapa --local-dir . --exclude "README.md" "meta.yaml" ".gitattributes"

./run.sh --skip_data_prep false --skip_train true
```

This will download the pre-trained model and run inference using the VoxLingua107 test data.

### Train and Evaluation Datasets

The model is evaluated on multiple language identification benchmarks with diverse characteristics:

| Dataset       | Domain      | #Langs. Train/Test | Dialect | Training Setup (VL107-only) |
| ------------- | ----------- | ------------------ | ------- | --------------------------- |
| [VoxLingua107](https://cs.taltech.ee/staff/tanel.alumae/data/voxlingua107/)  | YouTube     | 107/33             | No      | Seen                        |
| [Babel](https://citeseerx.ist.psu.edu/document?repid=rep1&type=pdf&doi=31a13cefb42647e924e0d2778d341decc44c40e9)         | Telephone   | 25/25              | No      | Unseen                      |
| [FLEURS](https://huggingface.co/datasets/google/xtreme_s)        | Read speech | 102/102            | No      | Unseen                      |
| [ML-SUPERB 2.0](https://huggingface.co/datasets/espnet/ml_superb_hf) | Mixed       | 137/(137, 8)       | Yes     | Unseen                      |
| [VoxPopuli](https://huggingface.co/datasets/facebook/voxpopuli)     | Parliament  | 16/16              | No      | Unseen                      |

### 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;">

| ESPnet Recipe                    | Config | VoxLingua107 | Babel | FLEURS | ML-SUPERB2.0 Dev | ML-SUPERB2.0 Dialect | VoxPopuli | Macro Avg. |
| ------------------------- | ----------- | ------------ | ----- | ------ | ---------------- | -------------------- | --------- | ---------- |
| <div class="hf-model-cell">[egs2/voxlingua107/lid1](https://github.com/espnet/espnet/tree/master/egs2/voxlingua107/lid1)</div> | <div class="config-cell">`conf/mms_ecapa_baseline`</div> | 94.2         | 86.7  | 95.8   | 89.0             | 73.4                 | 85.6      | 87.5       |

</div>


For more detailed inference results, please refer to the `exp_voxlingua107_raw/lid_mms_ecapa_baseline_raw/inference` directory in this repository.

## LID config

<details><summary>expand</summary>

```
config: /work/nvme/bbjs/qwang20/espnet/egs2/lid_delta/lid1/conf/mms_1b_ecapa/mms_ecapa_bs3min_baseline.yaml
print_config: false
log_level: INFO
drop_last_iter: false
dry_run: false
iterator_type: category
valid_iterator_type: category
output_dir: exp_voxlingua107_raw/lid_mms_ecapa_bs3min_baseline_delta_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: 30
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: 2
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: true
wandb_project: lid
wandb_id: null
wandb_entity: qingzhew-carnegie-mellon-university
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: 1000
batch_size: 20
valid_batch_size: null
batch_bins: 2880000
valid_batch_bins: null
category_sample_size: 10
train_shape_file:
- exp_voxlingua107_raw/lid_stats_16k/train/speech_shape
valid_shape_file:
- exp_voxlingua107_raw/lid_stats_16k/valid/speech_shape
batch_type: catpow
upsampling_factor: 0.5
language_upsampling_factor: 0.5
dataset_upsampling_factor: 0.5
dataset_scaling_factor: 1.2
max_batch_size: 16
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_voxlingua107/wav.scp
    - speech
    - sound
-   - dump/raw/train_voxlingua107/utt2lang
    - lid_labels
    - text
valid_data_path_and_name_and_type:
-   - dump/raw/dev_voxlingua107/wav.scp
    - speech
    - sound
-   - dump/raw/dev_voxlingua107/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: 5.0e-06
    betas:
    - 0.9
    - 0.98
scheduler: tristagelr
scheduler_conf:
    max_steps: 30000
    warmup_ratio: 0.3
    hold_ratio: 0.2
    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_voxlingua107/lang2utt
lang_num: 107
sample_rate: 16000
num_eval: 10
rir_scp: ''
model: espnet
model_conf:
    extract_feats_in_collect_stats: false
frontend: s3prl
frontend_conf:
    frontend_conf:
        upstream: hf_wav2vec2_custom
        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: rawnet3
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
loss: aamsoftmax_sc_topk
loss_conf:
    margin: 0.5
    scale: 30
    K: 3
    mp: 0.06
    k_top: 5
required:
- output_dir
version: '202412'
distributed: false
```

</details>



### Citation

```BibTex
@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}
}
```