GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
Finiflowlabs
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Run this model inference with full control and performance in your environment.
Model Details
Model Provider
Finiflowlabs
Model Tree
openai/whisper-medium
Input Modalities
Output Modalities
Supported Functionality
SAUTI ASR v1 is a Whisper-medium model fine-tuned on ~89 hours of Swahili speech data from the keystats/swahili_asr_data dataset. The model achieves 13.52% WER on a held-out test set, a 50% relative improvement over the zero-shot Whisper-medium baseline (~27% WER).
This model is part of the SAUTI platform — an African languages voice agent system combining ASR, LLM, and TTS for natural Swahili voice conversations.
| Parameter | Value |
|---|---|
| Base model | openai/whisper-medium (764M params) |
| Training data | keystats/swahili_asr_data (52,699 train / 5,271 val / 585 test) |
| Total audio | ~88.7 hours |
| Epochs | 5 |
| Effective batch size | 32 (16 per device × 2 gradient accumulation) |
| Learning rate | 1e-5 |
| Optimizer | AdamW |
| FP16 | Yes |
| Hardware | NVIDIA A100-SXM4-80GB |
| Training time | 7 hours 3 minutes |
| Steps | 8,235 |
| Metric | Value |
|---|---|
| Test WER | 13.52% |
| Test CER | 3.85% |
| Test Loss | 0.252 |
| Best Eval WER | 13.44% (step 7000) |
| Step | WER | Epoch |
|---|---|---|
| 500 | 27.23% | ~0.3 |
| 1000 | 21.37% | ~0.6 |
| 2500 | 15.58% | ~1.5 |
| 5000 | 14.11% | ~3.0 |
| 7000 | 13.44% | ~4.2 |
| 8235 (final) | 13.52% | 5.0 |
If you use this model, please cite: