khoben/whisper-ggml

Multilingual Whisper Base, Small and Turbo (GGML)

Three quantized, multilingual OpenAI Whisper models in the GGML file layout used by whisper.cpp, provided as standalone reusable speech-recognition artifacts. None is an English-only (.en) variant.

This repository redistributes unmodified upstream model files. Every model is a byte-for-byte copy of the file of the same name in ggerganov/whisper.cpp at revision 5359861c739e955e79d9a303bcbc70fb988958b1. The repository owner did not train or author the models, convert them, or re-quantize them. Each SHA-256 matches the upstream Hugging Face LFS object identifier at that revision.

Files

File	Whisper model	Quantization	Bytes	SHA-256
ggml-base-q5_1.bin	Base (multilingual)	q5_1	59707625	422f1ae452ade6f30a004d7e5c6a43195e4433bc370bf23fac9cc591f01a8898
ggml-small-q5_1.bin	Small (multilingual)	q5_1	190085487	ae85e4a935d7a567bd102fe55afc16bb595bdb618e11b2fc7591bc08120411bb
ggml-large-v3-turbo-q5_0.bin	Large v3 Turbo (multilingual)	q5_0	574041195	394221709cd5ad1f40c46e6031ca61bce88931e6e088c188294c6d5a55ffa7e2
ggml-base-q8_0.bin	Base (multilingual)	q8_0	81768585	c577b9a86e7e048a0b7eada054f4dd79a56bbfa911fbdacf900ac5b567cbb7d9
ggml-small-q8_0.bin	Small (multilingual)	q8_0	264464607	49c8fb02b65e6049d5fa6c04f81f53b867b5ec9540406812c643f177317f779f
ggml-large-v3-turbo-q8_0.bin	Large v3 Turbo (multilingual)	q8_0	874188075	317eb69c11673c9de1e1f0d459b253999804ec71ac4c23c17ecf5fbe24e259a1

The Q5 artifacts remain available for rollback. The Q8_0 files were copied from the same pinned upstream revision; downloaded bytes were independently hashed and their GGML headers identify Q8_0 (ftype 2007).

SHA256SUMS records the model and licensing-file digests. Pin a repository commit in download URLs (resolve/<commit>/<file>) and verify size and SHA-256 after download.

Provenance

• Original weights: OpenAI Whisper, MIT License, openai/whisper. The source revision recorded in NOTICE is 86098128c0b4f24f0e2aa2994de830614b474227; see LICENSE-openai-whisper.
• GGML conversion and quantization: the whisper.cpp project, MIT License, ggml-org/whisper.cpp; see LICENSE-whisper.cpp.
• Redistributed artifacts: ggerganov/whisper.cpp, revision 5359861c739e955e79d9a303bcbc70fb988958b1. Its model card declares license: mit.

Intended use

Multilingual speech-to-text research and inference with whisper.cpp or a compatible GGML loader. Choose a model and runtime configuration according to available memory, processing time and evaluation on the intended audio.

Limitations

• Whisper can omit, repeat or hallucinate words, particularly on noisy audio, music or short utterances. Performance varies by language, speaker and recording conditions.
• Model size, runtime cost and quality vary between Base, Small and Turbo; this repository makes no comparative accuracy or speed guarantee.
• Quantization can change outputs and recognition quality relative to unquantized weights.
• Whisper timestamps are estimates and should not be treated as independently verified forced-alignment boundaries.
• This redistribution includes no independent benchmark against reference transcripts.

Attribution and license

The models are distributed under the MIT License. Preserve the applicable copyright and license notices when redistributing them. See NOTICE, LICENSE-openai-whisper and LICENSE-whisper.cpp. The model binaries and licensing files are unchanged.