Research / Robust speech recognition and spoken search / [4]
Recognising rural Spanish with Radial LoRA: second place at ALBAYZIN 2026
Robust Speech Recognition under Spontaneous Rural Interviews via Radial Low Rank Adaptation: The IberSPEECH 2026 Speech-COSER Challenge. IberSPEECH 2026, ALBAYZIN Speech-COSER Challenge, 2026.
Adapting under 1 % of Whisper-large-v3's weights with R-LoRA on about 23 hours of rural interviews cuts the normalised word error rate on the blind test from 31.32 % to 26.31 %, and plain LoRA reaches only 29.24 %.
The problem
COSER, the Corpus Oral y Sonoro del Español Rural, collects ethnographic interviews recorded across rural Spain, with overlapping speech, background noise and, in some provinces, Galician, Catalan or Basque. Its transcripts are verbatim, so dialectal forms such as pos, pa or torcío are written as spoken. A recogniser trained on standard Spanish pushes them back to pues, para or torcido, and every such correction counts as an error.
What we did
We fine-tuned Whisper-large-v3 on about 23 hours of COSER audio with R-LoRA (Radial Low-Rank Adaptation), a method first introduced for clinical speech. R-LoRA is a variant of LoRA that tempers the size of each input before the low-rank update, with one learnable exponent per adapted layer: at 0 it is plain LoRA, at 1 the update sees only the direction of the input. It adds a single parameter per layer over LoRA. Because that tempering is non-linear, the adapters cannot be merged into the weights, so the model is packaged and served as one self-contained Hugging Face model.
Around it, a short inference stack: loudness normalisation, voice activity detection with Silero, and a language prior taken from the province of each recording.
What we found
On the blind test, R-LoRA reaches 26.31 % normalised WER, five points under the zero-shot model. Plain LoRA, which was only 0.33 points behind on development data, ends 2.93 points behind on the test set. LoRA was also more fragile in training: one checkpoint after its best it collapsed from 39.8 % to 55.6 % WER, while R-LoRA degraded gracefully.
A negative result worth knowing: every language-model post-corrector we tried (an n-gram rescorer, Gemma 3 12B with dialect-preservation instructions, and a fine-tuned mT5) made WER worse, by 0.2 to 2.9 points. They all pull the output towards standard Spanish, which a verbatim dialectal reference counts as wrong.
Why it matters
On a small in-domain dataset, the choice of adapter changed the test result more than the development data suggested. On the challenge’s final leaderboard the system finished second of six, 0.09 points behind the winner. The story of both tracks is in Two second places at ALBAYZIN 2026.
Cite
Ramírez-Sánchez J.M., Esperon-Couceiro A., Piñeiro-Martín A., Docío-Fernández L., García-Mateo C. Robust Speech Recognition under Spontaneous Rural Interviews via Radial Low Rank Adaptation: The IberSPEECH 2026 Speech-COSER Challenge. IberSPEECH 2026, ALBAYZIN Speech-COSER Challenge, 2026.