Downloads · 30 days
0
ICASSP2027_Jopara_ASR
victorshi119/ICASSP2027_Jopara_ASR
ICASSP2027_Jopara_ASR is a automatic speech recognition model from victorshi119. Use it when you need speech turned into text. The card lists the license as mit.
These are the adapter and router weights for the paper Jopara ASR: A Community-Verified Benchmark and Baselines for Spontaneous Spanish–Guaraní Code-Switching (Wenchen Shi, Shuju Shi; submitted to ICASSP 2027). Code,…
Access
Public
Updated Sep 24, 2026
Repo size
147 MB
Likes
0
Public
Repo makeup
Click a slice to open those files.
.pt147 MB · 100%
At a glance
- Task
- Automatic Speech Recognition
- License
- mit
- Access
- Public
- Created
- Sep 24, 2026
- Updated
- Sep 24, 2026
- SHA
- 0ecce2f3
Model card
From the Hugging Face model README
Jopara ASR: LoRA adapters and router weights
These are the adapter and router weights for the paper Jopara ASR: A Community-Verified Benchmark and Baselines for Spontaneous Spanish–Guaraní Code-Switching (Wenchen Shi, Shuju Shi; submitted to ICASSP 2027). Code, benchmark metadata, and full instructions are at https://github.com/victorshi119/ICASSP2027_Jopara_ASR.
All adapters are LoRA weights for the decoder's query and value projections of a frozen
OmniASR LLM 7B v2 (omniASR_LLM_7B_v2). They are not standalone models: load them with that base
model, using omnilingual-asr with our patch (see the GitHub repo's third_party/).
Files
| Path | Paper | Rank | Language mode | Trained on |
|---|---|---|---|---|
adapters/row5_spanish/ | Table 1 row 5 | 16 | spa_Latn | 1,083 Spanish-only utterances |
adapters/row6_guarani/ | Table 1 row 6 | 16 | grn_Latn | 2,569 Common Voice Guaraní clips + 19 Guaraní-only + 871 code-switching utterances |
adapters/row7_codeswitching/ | Table 1 row 7 | 32 | spa_Latn | 871 code-switching utterances |
adapters/moe_codeswitching_grn/ | Code-switching expert in row 8 | 32 | grn_Latn | 871 code-switching utterances |
adapters/moe_guarani_seed{13,42,123}/ | Guaraní expert in rows 8–9, one per seed | 16 | grn_Latn | same data as row 6 |
adapters/moe_null_codeswitching/ | Code-switching slot in row 9 | 8 | grn_Latn | none: all-zero weights, so the slot is the base model |
router/seed{13,42,123}/router_best.pt | Mixture-of-experts router, rows 8–9 | – | – | pooled encoder embeddings with soft targets |
Each adapter folder has lora_adapters.pt (a PyTorch state dict of the LoRA weights only) and
lora_config.json. The router architecture is defined in the GitHub repo
(training/moe_router/model/soft_regime_router.py, configured by
training/moe_router/configs/jmole_grn_bigred200.yaml). The Spanish expert in rows 8–9 is
row5_spanish. SHA256SUMS lists a checksum for every file.
Results
Results on the 150-utterance test set, as CER / WER (%):
| System | CER / WER |
|---|---|
| OmniASR, auto-detect (no adapter) | 28.4 / 50.4 |
| OmniASR, Guaraní language mode (no adapter) | 25.0 / 47.9 |
| Row 5, Spanish adapter | 37.7 / 63.2 |
| Row 6, Guaraní adapter | 29.1 / 53.8 |
| Row 7, code-switching adapter | 29.7 / 52.9 |
| Row 8, mixture of experts (3 seeds) | 28.8±2.0 / 50.9±2.8 |
| Row 9, mixture of experts with an empty code-switching slot (3 seeds) | 24.3±0.1 / 46.1±0.2 |
None of the adapters or routers significantly beats simply running the base model in the Guaraní language mode. See the paper for confidence intervals.
Training data
The training data comes from CEGPA (California Language Archive, doi:10.7297/X2CC0ZW0) and Common Voice Guaraní. Neither is redistributed here or in the GitHub repo.
License
The adapter and router weights are released under MIT. The base model, OmniASR, is governed by its own license.
- Task
- Automatic Speech Recognition
- License
- mit
- Languages
- gn, es
- Created
- Sep 24, 2026
- Updated
- Sep 24, 2026