Step-Audio 2 Mini Luganda-English Bidirectional S2ST
This repository contains a full merged model for bidirectional Luganda-English
speech-to-speech translation. It was created by merging the LoRA adapter yigagilbert/stepaudio2-mini-luganda-english-bidirectional-s2st-lora into
stepfun-ai/Step-Audio-2-mini.
The separate adapter-only repository should remain available for users who prefer PEFT loading or want the smaller adapter artifact. This full repository is intended for simpler deployment and inference where loading a single model repo is preferable.
Intended Use
Research and development for Luganda-English speech translation in both directions:
- Luganda speech input -> English text and English speech tokens
- English speech input -> Luganda text and Luganda speech tokens
Validate outputs with native speakers before production or high-stakes use.
Source Model and Adapter
- Base model:
stepfun-ai/Step-Audio-2-mini - LoRA adapter:
yigagilbert/stepaudio2-mini-luganda-english-bidirectional-s2st-lora - Merge script:
scripts/push_full_model_to_hub.py
Training Data
The model was trained from yigagilbert/luganda-english-cleaned-v1-split by expanding
each aligned speech pair into two direction-conditioned supervised examples. Each example
uses a direction-specific system prompt so that one adapter can learn both translation
directions.
Evaluation Summary
The checkpoint used for this merge was evaluated on 200 held-out validation examples per direction. The metrics below were generated with the adapter-loaded fine-tuned model before merge; the merged model contains the same adapted weights and is expected to match these results aside from normal deterministic or runtime differences. Re-run evaluation directly on this repository before a strict release if exact reproducibility is required.
Text Metrics
| Direction | System | Count | BLEU higher | chrF higher | WER lower | COMET higher |
|---|---|---|---|---|---|---|
| Luganda -> English | Focused one-direction LoRA | 200 | 32.530 | 54.535 | 0.574 | 0.717 |
| Luganda -> English | This bidirectional model | 200 | 19.032 | 42.447 | 0.737 | 0.640 |
| English -> Luganda | Focused one-direction LoRA control | 200 | 0.120 | 14.050 | 1.300 | - |
| English -> Luganda | This bidirectional model | 200 | 11.155 | 45.331 | 0.811 | 0.678 |
| English -> Luganda | ASR + MT cascade baseline | 200 | 4.026 | 34.734 | 2.039 | 0.624 |
Speech Metrics
Speech metrics were computed on 20 audio-aligned samples per direction using WavLM-large SpeechBERTScore-style similarity and MFCC+DTW MCD. These are supporting audio-channel diagnostics, not standalone translation-quality measures.
| Direction | Count | SpeechBERT P higher | SpeechBERT R higher | SpeechBERT F1 higher | MCD lower |
|---|---|---|---|---|---|
| Luganda -> English | 20 | 0.560 | 0.582 | 0.570 | 639.360 |
| English -> Luganda | 20 | 0.502 | 0.572 | 0.523 | 614.130 |
Retention and Composition Probes
The bidirectional model does not recover unrelated multilingual speech translation ability on small FLEURS probes: Spanish -> English BLEU is 0.293 and Japanese -> English BLEU is 0.163 on 20 samples each. However, it does recover target-script compliance on Luganda -> Chinese and Luganda -> Japanese prompt probes (100% target-script hit rate in both cases). The semantic quality of those composition outputs is poor, with roundtrip BLEU below 1, so these probes should not be read as successful Luganda -> Chinese or Luganda -> Japanese translation.
Interpretation
This model demonstrates that one direction-conditioned LoRA adaptation of Step-Audio-2-mini can learn both Luganda -> English and English -> Luganda speech translation. Compared with the earlier Luganda-to-English-only adapter, it adds English-to-Luganda capability but is weaker on Luganda-to-English. The bidirectional run was shorter (9,000 steps) than the focused one-direction run, so the Luganda-to-English gap combines multi-task interference and training duration.
Notes
If this repository includes token2wav/, those assets are provided to support waveform
synthesis from generated audio tokens. Some inference clients may still use the official
Step-Audio2 runtime code for token-to-waveform conversion.
License
The training code and adapter metadata are Apache-2.0. Because this merged repository contains base-model weights, users must also comply with the base model license and any dataset licensing constraints.