Addis Scribe Streaming: Afaan Oromo
Realtime Afaan Oromo speech recognition by Addis AI. Text appears while people speak and is final about
40 ms after they stop. Built on NVIDIA Nemotron 3.5 ASR Streaming (0.6B) and trained by Addis AI on Afaan Oromo
speech. The same model runs Addis Scribe's language=om in the Addis AI API.
Benchmark: 150 WAXAL Afaan Oromo test clips
23 speakers, 45 minutes, human transcripts; none of it was used in training. One scorer for every system; WER = word error rate, CER = character error rate, lower is better.
| System | Mode | WER | CER | Text finished after you stop | First words on screen |
|---|---|---|---|---|---|
| Addis Scribe (this model) | live, 1.12 s chunks | 21.1% | 6.8% | 43 ms | 2.21 s |
| Addis Scribe (this model) | live, 0.32 s chunks | 21.5% | 6.9% | 37 ms | 1.57 s |
| Addis Scribe (this model) | whole file | 22.3% | 7.6% | – | – |
| Ethio-ASR Oromo | offline | 24.8% | 7.1% | – | – |
| Ethio-ASR multilingual 600M | offline | 25.7% | 7.2% | – | – |
| WAXAL Whisper-small Oromo | offline | 28.2% | 8.2% | – | – |
| WAXAL MMS-300M Oromo | offline | 29.5% | 7.9% | – | – |
| Google Gemini Flash (latest) | whole file | 29.8% | 9.7% | – | – |
| Google Gemini 4 Argon | whole file | 29.9% | 7.8% | – | – |
| Google Chirp 2 (om-ET) | batch | 34.8% | 9.9% | – | – |
| Google Chirp 2 (om-ET) | live | 42.4% | 14.2% | 1,051 ms | 6.37 s |
Latency for this model is measured with audio fed at real-time speed and no network hop; Google's figures include the network. Every clip, reference, transcript, the scorer and scripts to rerun each system with your own keys are in the benchmark folder: addisai/addis-scribe-benchmark.
Use it
import nemo.collections.asr as nemo_asr
model = nemo_asr.models.ASRModel.restore_from("addis-scribe-oromo-streaming.nemo")
model.set_inference_prompt("om-ET") # the Afaan Oromo prompt (index 59)
print(model.transcribe(["speech_16k.wav"])[0])
For live use, run it as a cache-aware streaming model with 0.32 s or 1.12 s chunks, the same way as
Addis Scribe Streaming (Amharic) and its stream_infer.py,
with the prompt om-ET / index 59. Requires NVIDIA NeMo 3.0. Or call it through the
Addis AI API with language=om, including speaker labels, word timestamps and captions.
Model
| Architecture | Cache-aware FastConformer encoder, hybrid RNNT + CTC, 0.6B parameters |
| Decoding | Greedy RNNT; language prompt om-ET (index 59) |
| Chunk sizes | 0.32 s or 1.12 s |
| Input / output | 16 kHz mono audio / Afaan Oromo text in Qubee, lowercase, no punctuation (the apostrophe for the glottal stop is kept) |
| Files | addis-scribe-oromo-streaming.nemo, SHA256SUMS |
Limits
- No punctuation or capitals.
- Afaan Oromo spelling varies (vowel length, doubled consonants, the glottal mark); some measured errors are spelling differences from the reference rather than mishearing.
- Tested on Afaan Oromo only.
License
OpenMDW 1.1, the same as the base model's terms allow. Made by Addis AI.