This variant uses a broader Amharic speech-text mixture and is the recommended release when generalization beyond the curated training domain is the priority.

Benchmark

Results use fixed held-out evaluation sets and identical deterministic generation settings (language=None, max_new_tokens=512, batch size 1). Lower WER and CER are better.

Evaluation	Curated variant	Broader mixed variant
Curated-source heldout WER	36.99%	37.46%
Curated-source heldout CER	15.71%	16.18%
WAXAL WER	37.56%	36.57%
WAXAL CER	16.45%	15.57%
Supplemental-source heldout WER	Not evaluated	31.47%
Supplemental-source heldout CER	Not evaluated	13.87%

The broader mixed variant improves WAXAL WER by 0.99 percentage points over the curated variant, while the curated variant remains 0.47 points better on its source-matched heldout.

Live Colab Demo

Open the Gradio demo notebook, then open it in Google Colab and run the cells in order. The final cell creates a temporary public Gradio link while the notebook remains connected.

Usage

pip install -U qwen-asr

python

import torch
from qwen_asr import Qwen3ASRModel

model = Qwen3ASRModel.from_pretrained(
    "b1n1yam/qwen3-asr-0.6b-amharic-mixed",
    dtype=torch.bfloat16,
    device_map="cuda:0",
    max_inference_batch_size=8,
    max_new_tokens=512,
)

result = model.transcribe(
    audio="path/to/audio.wav",
    language=None,
)

print(result[0].text)

Audio should contain clear Amharic speech. Keep language=None; this is the setting used for the reported benchmarks.

Intended Use

General-purpose Amharic speech transcription
Research and evaluation of Amharic ASR
Further domain adaptation with appropriately licensed data

Limitations

Performance varies by recording conditions, speaker, dialect, domain, and audio duration.
The model can omit words or generate incorrect text. Human review is required for consequential uses.
The benchmark covers a limited set of held-out corpora and should not be interpreted as universal Amharic ASR performance.
Supplemental training material received structural validation and exact audio deduplication, but was not assigned the same row-level quality tiers as the curated corpus.
This release is not designed for speaker identification, translation, or timestamp alignment.

Training Data Disclosure

The model was adapted with a privately maintained mixture of curated and supplemental Amharic speech-text data. Source manifests and row-level metadata are not included in this repository.

License and Attribution

This model is released under the Apache-2.0 license, following the upstream Qwen3-ASR-0.6B license. Please also review the upstream model card.