Addis Scribe Streaming

Realtime Amharic speech recognition by Addis AI. Text appears while people speak, in 0.32-second steps, and is final about 30 ms after they stop. Built on NVIDIA Nemotron 3.5 ASR Streaming (0.6B), trained on 5,000+ hours of quality-checked Amharic speech.

	Addis Scribe Streaming	Google Chirp 3 (streaming)
Word error rate, natural speech	28.4%	32.5%
Character error rate, natural speech	13.1%	16.5%
Final text after the speaker stops	30 ms	1,411 ms
First words on screen	1.56 s	6.67 s

Streaming support

"Streaming" here means the model reads each new piece of audio once, keeps what it learned in a cache, and updates the text without re-reading earlier audio. Cost and delay stay flat however long someone talks.

System	Streaming	How it handles live audio
Addis Scribe Streaming	Native	Cache-aware: each 0.32 s (or 1.12 s) chunk is processed once.
Google Chirp 3	Native (cloud API)	Server-side streaming over the network.
Shook (Whisper medium)	Not supported	Whisper models transcribe complete audio, not a live stream. Our benchmark runs it live using the method of WhisperLiveKit, a widely used open-source tool for live Whisper: once per second it transcribes all audio received so far and displays a word once two consecutive transcriptions contain it. Because each pass covers the whole recording so far, the delay grows the longer someone speaks: measured median 10.8 s after speech ends.
Hohe (wav2vec2-BERT)	No	Offline only: needs the whole recording before it can answer.

Benchmarks

Every system is scored with the same text normaliser (Unicode NFC, punctuation and symbols removed). WER = word error rate, CER = character error rate; lower is better.

Natural speech: 100 WAXAL Amharic test clips

22 speakers, 29 minutes, human transcripts. None of it was used in training. Clips were picked at random (fixed seed) from WAXAL amh test, 2 to 30 s long.

System	Mode	WER	CER	Final text after speech ends	First words
Addis Scribe Streaming	streaming 0.32 s	28.4%	13.1%	30 ms	1.56 s
Addis Scribe Streaming	streaming 1.12 s	27.3%	12.4%	34 ms	2.21 s
Google Chirp 3 (am-ET)	streaming, realtime	32.5%	16.5%	1,411 ms	6.67 s
Google Chirp 3 (am-ET)	batch, whole file	29.2%	13.5%	n/a	n/a
Shook (Whisper medium)	live, WhisperLiveKit method (1 s updates)	35.0%	16.6%	10,839 ms	3.66 s
Shook (Whisper medium)	offline	35.0%	16.6%	n/a	n/a
Hohe (wav2vec2-BERT)	offline only	25.4%	11.2%	n/a	n/a

Latency for Addis Scribe and Shook is measured on one NVIDIA A100 with no network hop, with audio fed at real-time speed; Google's figures include the network round trip to its EU endpoint from the benchmark machine.

FLEURS Amharic test (read sentences)

System	Mode	WER	CER
Addis Scribe Streaming	streaming 0.32 s (first 150 clips)	19.8%	8.2%
Hohe	offline (same 150 clips)	20.6%	7.9%
Addis Scribe Streaming	offline (all 516 clips)	20.6%	7.7%
Hohe	offline (all 516 clips)	20.2%	7.3%
Previous Addis AI Nemotron	streaming 0.32 s (first 150 clips)	35.0%	17.0%

Every per-clip transcript, latency and the scoring script are in benchmark/, so these numbers can be reproduced exactly.

Quickstart

from stream_infer import AmharicStreamer

scribe = AmharicStreamer("addis-scribe-streaming.nemo", chunk="320ms")   # or "1120ms"
for frame in microphone_frames():        # float32 mono, 16 kHz, any frame size
    print(scribe.feed(frame))            # transcript so far
final_text = scribe.finish()             # at end of speech; scribe.reset() for the next utterance

Command line: python stream_infer.py addis-scribe-streaming.nemo speech_16k.wav [320ms|1120ms]. Requires NVIDIA NeMo 3.0 with numba-cuda[cu12], numpy<2.4 and nvidia-nvjitlink-cu12>=12.9. Fed 100 ms frames like a live microphone, AmharicStreamer reproduces the reference streaming transcripts exactly (100 of 100 benchmark clips).

Model


Architecture	Cache-aware FastConformer encoder, hybrid RNNT + CTC, 0.6B parameters
Decoding	Greedy RNNT; language prompt `am-ET` (index 49)
Chunk sizes	0.32 s (default) or 1.12 s
Input / output	16 kHz mono audio / Amharic text without punctuation
Files	`addis-scribe-streaming.nemo`, `serving.json`, `stream_infer.py`, `SHA256SUMS`

Training

Trained on 5,000+ hours of quality-checked Amharic speech, then refined on human-transcribed speech. FLEURS, WAXAL test and the evaluation speakers were never used in training.

Limits

On natural conversational speech it is about 3 WER points behind Hohe, which is offline only.
No punctuation (training text had punctuation removed).
Tested on Amharic only; other languages are out of scope.

References


Base model	NVIDIA Nemotron 3.5 ASR Streaming 0.6B
Toolkit	NVIDIA NeMo
Google Chirp 3	Chirp 3 model documentation
Hohe	snapwre/hohe-asr-amharic
Shook	b1n1yam/shook-medium-amharic-2k
WhisperLiveKit	QuentinFuxa/WhisperLiveKit
WAXAL (natural-speech test set)	google/WaxalNLP, `amh` test split, CC-BY-SA-4.0
FLEURS (read-speech test set)	google/fleurs, `am_et` test split

addisai/addis-scribe-streaming