Addis Scribe now transcribes Wolaytta, live and from uploads. We compared it with the Ethio-ASR models and Google Gemini on the same held-out natural speech, with the same scorer for every system. The test set is small, so the margins here are narrow.
Addis Scribe has the lowest word error rate of the systems here, by a narrow margin on a small set, and it is the only one that transcribes live.
On 103 WAXAL test clips, streaming in 1.12 s steps, Addis Scribe records 37.4% WER and 14.6% CER. Ethio-ASR Wolaytta records 39.4% and Ethio-ASR multilingual 600M 40.3%, offline. On character error rate, both Ethio-ASR models are slightly lower (13.8% and 14.3%). With 103 clips, a point or two either way is within noise.
Read the full accuracy tablesGiven the whole clip, Google Gemini Flash (latest) records 90.5% WER and 52.6% CER.
Read the full accuracy tables103 clips from the WAXAL Wolaytta test split, natural speech with human transcripts. Every system transcribed the same clips.
WAXAL Wolaytta, 103 clips, lower is better
WAXAL Wolaytta, 103 clips, lower is better
| System | Mode | WER | CER |
|---|---|---|---|
| Addis Scribe Wolaytta | Live, 1.12 s chunks | 37.4% | 14.6% |
| Addis Scribe Wolaytta | Live, 0.32 s chunks | 38.5% | 14.6% |
| Addis Scribe Wolaytta | Whole file | 39.1% | 15.5% |
| Ethio-ASR Wolaytta | Offline | 39.4% | 13.8% |
| Ethio-ASR multilingual 600M | Offline | 40.3% | 14.3% |
| Google Gemini Flash (latest) | Whole file | 90.5% | 52.6% |
Bold marks the best value in each column. The charts cut the Gemini bars; their values are in the table.
Addis Scribe streams live audio. The other systems were given the whole recording, so they are scored for accuracy only.
| System | Live audio | How it was run |
|---|---|---|
| Addis Scribe Wolaytta | Native streaming | Streams in 0.32 s or 1.12 s chunks through the Addis Scribe API, so text appears while people speak. |
| Ethio-ASR Wolaytta | Offline only | Needs the whole recording before it can answer. Scored for accuracy. |
| Ethio-ASR multilingual 600M | Offline only | Needs the whole recording before it can answer. Scored for accuracy. |
| Google Gemini Flash (latest) | Offline only | Given the whole clip with a transcription prompt. No live mode, so scored for accuracy. |
The test set is public. None of these test clips were used to train Addis Scribe.
| Test set | Where to find it | Clips | What it measures |
|---|---|---|---|
| WAXAL | google/WaxalNLP | Wolaytta test split, 103 of its 3,014 clips | Natural speech with human transcripts |
Every system on a set is scored with the same scorer and text normaliser, against the same human transcripts. WER is word error rate and CER is character error rate, both lower is better.
Add language=wal to a Scribe upload, live session or long-file job. Uploads, live streaming, word timestamps, captions and speaker labels all work in Wolaytta, at the same per-character price as Amharic. Wolaytta text comes back in lowercase without punctuation; an apostrophe marks the glottal stop.