Step-Audio 2 Mini Luganda-to-English S2ST LoRA
This repository contains a LoRA adapter for stepfun-ai/Step-Audio-2-mini trained for
Luganda speech input to English speech output. The adapter requires the base model and
Step-Audio 2 token2wav assets for waveform synthesis.
Adapter repository: yigagilbert/stepaudio2-mini-luganda-english-s2st-lora
Intended use
Research and development for Luganda-to-English speech translation. Validate with native speakers before production use.
Training data
Configured for yigagilbert/luganda-english-cleaned-v1-split with columns:
audio_lug, audio_eng, text_lug, text_eng, id, src_dur_s, tgt_dur_s,
dur_ratio, src_speech_ratio, tgt_speech_ratio.
Evaluation
The table below summarizes the held-out validation evaluation used during development. All text metrics were computed on 200 validation examples. Speech metrics were computed on the aligned 197-example subset for which generated audio existed.
The merged full-model repository contains the same adapted weights as this adapter after LoRA merge, so it is expected to match these results aside from normal deterministic or runtime differences. Re-run evaluation directly on the merged repository before a strict release if exact reproducibility is required.
Text and Semantic Metrics
| System | Loading Form | Count | BLEU higher | chrF higher | WER lower | COMET higher | BLASER ref higher | BLASER QE higher |
|---|---|---|---|---|---|---|---|---|
| Base Step-Audio-2-mini | Base only | 200 | 0.012 | 5.152 | 10.702 | 0.386 | 1.713 | 2.164 |
| Fine-tuned Step-Audio | Base + this LoRA adapter | 200 | 32.530 | 54.535 | 0.574 | 0.717 | 3.762 | 3.723 |
| Full merged fine-tuned model | Same adapted weights, merged | 200 | 32.530* | 54.535* | 0.574* | 0.717* | 3.762* | 3.723* |
| Cascade baseline | ASR + MT + TTS | 200 | 36.778 | 57.971 | 0.521 | 0.737 | 3.839 | 3.776 |
* The merged full-model row reflects the adapter evaluation because the merged
repository is produced by folding this LoRA adapter into the same base model.
Speech Metrics
| System | Count | chrF higher | SpeechBERT P higher | SpeechBERT R higher | SpeechBERT F1 higher | MCD lower |
|---|---|---|---|---|---|---|
| Fine-tuned Step-Audio | 197 | 54.582 | 0.644 | 0.648 | 0.645 | 629.718 |
| Full merged fine-tuned model | 197 | 54.582* | 0.644* | 0.648* | 0.645* | 629.718* |
| Cascade baseline | 197 | 57.893 | 0.603 | 0.622 | 0.612 | 613.212 |
The unfine-tuned base model emitted no valid speech-token sequences in this 200-sample run, so SpeechBERTScore and MCD were not computed for it.
Interpretation
The LoRA adapter changes the base model from an unusable zero-shot system into a functioning Luganda-to-English speech translation model. The cascade remains stronger on text and semantic metrics, while the fine-tuned Step-Audio system is operationally simpler and scored higher on the WavLM-based SpeechBERTScore F1 proxy.
License
Adapter code and metadata are Apache-2.0. Check the dataset license separately before redistribution.