File size: 2,990 Bytes
faf1b07
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
db60d7a
faf1b07
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
db60d7a
faf1b07
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
---
license: mit
library_name: pytorch
tags:
  - speaker-recognition
  - speaker-encoding
  - speech
  - indic
  - cross-lingual
  - voice-cloning
language:
  - en
  - hi
  - te
  - ta
---

# LASE r1 — Language-Adversarial Speaker Encoder

Reference checkpoint for the paper *"LASE: Language-Adversarial Speaker Encoding for Indic Cross-Script Identity Preservation"* ([arXiv:2605.00777](https://arxiv.org/abs/2605.00777)).

LASE is a 256-d speaker embedding that preserves speaker identity across Devanagari, Telugu, Tamil, and Latin scripts. It wraps a frozen `microsoft/wavlm-base-plus` backbone with a 2-layer projection MLP and a gradient-reversal language classifier (~170k trainable params).

## Headline result

| Encoder | Western voices gap | Indian voices gap |
|---|---|---|
| WavLM-base-plus-sv (off-the-shelf) | 0.082 | 0.006 |
| ECAPA-TDNN (off-the-shelf) | 0.105 | 0.058 |
| ECAPA + GRL (ablation) | 0.027 | 0.037 |
| **LASE r1 (ours)** | **0.013** | **−0.000** |

Lower is better. *gap* = within-script median minus cross-script median for the same speaker. LASE r1's bootstrap 95% CI on gap straddles zero on both held-out corpora.

## Usage

```python
from huggingface_hub import hf_hub_download
import torch
# clone github.com/praxelhq/lase first for the model code
from models.lase import LASE, LambdaSchedule, WavLMSpeakerEncoder

ckpt_path = hf_hub_download("Praxel/lase-r1", "last.pt")
backbone = WavLMSpeakerEncoder("microsoft/wavlm-base-plus", embedding_dim=256, freeze_backbone=True)
model = LASE(backbone, embedding_dim=256, n_languages=4,
             lambda_schedule=LambdaSchedule(200, 500, 0.1))
model.load_state_dict(torch.load(ckpt_path)["model"], strict=False)
model.eval()

# wav: (B, T) float32 at 16 kHz, ~2 seconds
embedding = model(wav)["embedding"]   # (B, 256)
```

## Training

- **Backbone**: `microsoft/wavlm-base-plus` (frozen)
- **Projection MLP**: 768 → 512 → 256 (~170k params)
- **Losses**: SupCon (voice identity) + GRL CE (4-language adversarial)
- **λ schedule**: warmup 0 for 200 steps, ramp to 0.1 over 500 steps, hold
- **Optimisation**: 1000 steps, batch 16, AdamW, LR 1e-4
- **Data**: 1118 same-voice cross-script pairs from 8 ElevenLabs Multilingual voices, gated through WavLM-cosine ≥ 0.90
- **Hardware**: 1× A10G on Modal, ~17 min, ~$0.31

## Datasets

- Training: [`Praxel/codeswitch-pairs-lase`](https://huggingface.co/datasets/Praxel/codeswitch-pairs-lase)
- Western held-out: [`Praxel/codeswitch-pairs-lase-heldout`](https://huggingface.co/datasets/Praxel/codeswitch-pairs-lase-heldout)
- Indian held-out: [`Praxel/codeswitch-pairs-lase-indian`](https://huggingface.co/datasets/Praxel/codeswitch-pairs-lase-indian)

## License

MIT.

## Citation

```bibtex
@misc{lase2026,
  title={{LASE}: Language-Adversarial Speaker Encoding for {Indic} Cross-Script Identity Preservation},
  author={Menta, Venkata Pushpak Teja},
  year={2026},
  eprint={2605.00777},
  archivePrefix={arXiv},
  primaryClass={eess.AS},
}
```