The study reports a word error rate (WER) of 37.5% and a character error rate (CER) of 7.45% after adapting Whisper Small to Baniwa. The authors present the figures as an initial Baniwa automatic speech-recognition baseline and a foundation for future research.
Those scores came from 1,373 manually transcribed audio recordings totaling approximately 0.54 hours, or 32.4 minutes. Most of the material consisted of elicited isolated words and short utterances, defining the setting represented by the reported numbers.
How the baseline was built
The study evaluates whether Whisper can be adapted to Baniwa through supervised fine-tuning on a manually transcribed speech corpus when training data are limited. The recordings were randomly divided into approximately 90% training, 5% validation and 5% test subsets.
Before training, the recordings were converted to 16 kHz audio, represented as log-Mel spectrograms—a numerical description of sound across time and frequency—and paired with tokenized transcriptions. Because Baniwa was not officially supported among Whisper’s predefined language tokens, Spanish transcription prompts were used. No post-processing was used.
The fine-tuning run used a learning rate of 1 × 10−5, batch size 8, gradient accumulation of 2, a maximum of 300 steps and FP16 mixed precision. Checkpoints were evaluated and saved every 100 optimization steps, with the best checkpoint selected using validation performance.
The two measures told different stories
Across the three recorded checkpoints, WER was 55.00% at step 100, 37.50% at step 200 and 40.00% at step 300. CER was 10.05%, 7.45% and 7.28% at those same steps. WER therefore reached its lowest reported value at step 200, while CER reached its lowest at step 300.
That mismatch qualifies the headline figures: the supplied analysis does not fully clarify whether the 37.5% WER and 7.45% CER came from one jointly selected checkpoint or from metric-specific minima.
Meanwhile, training loss fell from 1.0883 at step 100 to 0.0669 at step 300, while validation loss fell from 0.4313 to 0.3652. The document says longer schedules showed signs of overfitting, which is why the final setting used 300 optimization steps.
A result bounded by the dataset
The authors present the work as an initial Baniwa ASR baseline and a foundation for future research. The corpus description places that baseline in a tightly defined setting: 1,373 recordings and 32.4 minutes of predominantly elicited isolated words and short utterances.
The setup also leaves several possible improvements untested. No external language model, pronunciation lexicon, data augmentation or post-processing was used, so the reported figures describe this configuration rather than a comparison of those alternatives.
Access and status
The corpus is not publicly available because of ownership and community access considerations. Access may be granted on reasonable request and with authorization from the data custodians, while source code for model training and evaluation is available from the corresponding author on reasonable request.
The report is a version 1 arXiv preprint dated 26 Aug 2026. The authors report no specific funding, declare no conflict of interest, and list ethics approval and consent items as not applicable.
Paper data and sources
Original title: Fine-Tuning Whisper for Automatic Speech Recognition in Baniwa: A Preliminary Study
Authors: Leonardo Duart, Tiago Fonseca, Thiago Chacón
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text