whisper-ru GGML artifacts
GGML artifacts for running
dvislobokov/whisper-large-v3-turbo-russian
with whisper.cpp in the whisper-ru local ASR server.
Files
| File | Purpose |
|---|---|
ggml-large-v3-turbo-russian-q5_k.bin | Recommended runtime model; Q5_K quantized GGML |
ggml-large-v3-turbo-russian-f16.bin | Optional conversion artifact; useful for re-quantization |
For public distribution, publishing only the Q5_K file is enough for normal API users. Keep F16 private unless you explicitly want to support downstream re-quantization.
Source model
- Base/fine-tuned model:
dvislobokov/whisper-large-v3-turbo-russian - Declared license of the source model card: MIT
- Base model lineage:
openai/whisper-large-v3-turbo
Usage
The code repo's idempotent setup command downloads this ready-to-run Q5_K file by default:
docker compose run --rm whisper-api prepare-model
To rebuild from the source Hugging Face model instead:
PREPARE_MODEL_FROM_SOURCE=true docker compose run --rm whisper-api prepare-model
Source rebuild pipeline:
- Download the Hugging Face model snapshot.
- Convert Transformers/Safetensors weights to GGML F16 with whisper.cpp.
- Quantize F16 to Q5_K with
whisper-quantize.
Runtime behavior
The whisper-ru API server uses Silero VAD by default before decoding. This
prevents non-speech audio, silence, and microphone tails from reaching Whisper
and producing subtitle-like hallucinations such as Продолжение следует....
Clients can explicitly request raw Whisper decoding with vad=false or
X-VAD: false.
Reference benchmark
AMD Radeon 780M / Vulkan / Q5_K:
Audio duration: 60.000s
Wall time: 3.534s
RTF: 0.059
Benchmarks are hardware- and driver-dependent.