Open source · Apache-2.0 · CPU only

Belarusian speech recognition that runs on a CPU

NVIDIA's Belarusian FastConformer, packaged for ONNX Runtime, with a server that speaks OpenAI's and whisper.cpp's APIs. Measured on FLEURS against Whisper large-v3.

Беларускае распазнаванне маўлення на звычайным працэсары.

How accurate

Benchmark

CER and WER are corpus rates after lower-casing and removing punctuation. belarusian-asr writes numbers as digits, like FLEURS mostly does; "no digits" shows the sentences whose reference has no digits, where a number written differently cannot count against either system.

Listen and compare

The reference and belarusian-asr are Belarusian. Whisper large-v3 is shown exactly as it answered, and that is the problem this project solves: it mixes Russian and Ukrainian into Belarusian ("приправляна", "країна", "же" instead of "жа"). Letters that do not exist in Belarusian are boxed.

~~struck~~ in the reference: missedmarked in an output: not in the referenceи ї not a Belarusian letter (Russian or Ukrainian)

Use it

pip install "belarusian-asr[server] @ git+https://github.com/YauhenBichel/belarusian-asr"
belarusian-asr transcribe clip.wav
belarusian-asr serve --port 11805      # POST /v1/audio/transcriptions, POST /inference