alekringtonnn-ai/zubr-whisper

🦬🔊 Zubr Whisper

Repository: alekringtonnn-ai/zubr-whisper

—

📖 What This Model Is

zubr-whisper is the GGUF version of the audio model by alekringtonnn-ai for the task of Automatic Speech Recognition (ASR) — speech-to-text transcription. The model converts an audio file → text.

The model is built on openai/whisper-small — a compact multilingual version of OpenAI’s Whisper (~244 Million parameters). Whisper is the most popular open-source speech recognition engine, supporting ~99 languages, including Russian.

The GGUF format is a universal format for quantized weights: the model runs on any platform — Windows, Linux, macOS, Android — on a regular CPU, without expensive GPUs.

💡 For Macs with Apple Silicon (M1–M4), there is a separate MLX version of this same model: alekringtonnn-ai/zubr-whisper-mlx.

Key Features

• Speech Recognition (ASR): the model’s task is audio transcription (speech → text), not chat dialogue.
• Base — Whisper-small from OpenAI: a proven architecture with multilingual support (~99 languages, including Russian).
• Cross-platform: GGUF + whisper.cpp = runs on CPU and GPU (NVIDIA, AMD, Intel) on Windows, Linux, and macOS.
• Compactness: the versions weigh from ~90 MB (Q2) to ~259 MB (Q8) — the model runs even on low-power laptops.
• Subtitles and timecodes: like Whisper, the model can produce per-segment transcripts with timestamps.

⚠️ Important: this is an audio transcription model — it is not intended for text generation from prompts (chat), image processing, etc. For the conversational models in the Zubr lineup, see zubr-mini-1.9-3b.

Model Tree

openai/whisper-small                          ← base model (OpenAI)
        │ (GGUF quantization by alekringtonnn-ai)
alekringtonnn-ai/zubr-whisper                 ← GGUF version (THIS repository) — whisper.cpp, all platforms
        │ (conversion to MLX)
alekringtonnn-ai/zubr-whisper-mlx             ← MLX version for Apple Silicon

—

📊 Versions (Quantizations) in the Repository

The repository offers 6 GGUF quantization options — from an ultra-compact 2-bit one to a nearly original 8-bit one:

Version	File	Size	What it is	What it’s for
Q2_K	q2_k.gguf	~89.7 MB	The most aggressive 2-bit quantization	Devices with minimal memory (smartphones, single-board computers). Recognition quality suffers noticeably
Q4_0	whisper-small-q4_0.gguf	~139 MB	Classic 4-bit quantization (fast, simple format)	A well-established legacy format: high speed, minimal requirements
Q4_1	whisper-small-q4_1.gguf	~154 MB	4-bit quantization with an accuracy layer (better than Q4_0)	Slightly higher quality than Q4_0 at almost the same speed
Q5_0	whisper-small-q5_0.gguf	~175 MB	5-bit quantization — a high quality level	A good accuracy/size balance for regular use
Q5_1 ⭐	whisper-small-q5_1.gguf	~190 MB	5-bit quantization with an accuracy layer	The recommended choice: quality close to Q8 at a noticeably smaller size
Q8_0	whisper-small-q8_0.gguf	~259 MB	8-bit quantization — virtually lossless quality	Maximum recognition quality; the file is still compact

💡 How to choose: for transcription, quality matters more than size — go with Q5_1 (recommended) or Q8_0 (maximum). Q4_0/Q4_1 — if speed on weak hardware matters more; Q2_K — only for very limited devices.

—

📥 How to Download

Requirements

• Any modern CPU (x86 with AVX/AVX2 or ARM); a GPU is optional (NVIDIA/AMD/Intel accelerate recognition).
• RAM: from ~512 MB free for Q2 up to ~1 GB for Q8.
• Free disk space: 90–260 MB depending on the version.
• To run — whisper.cpp installed.

Method 1: huggingface-cli (recommended)

# Install the CLI
pip install -U «huggingface_hub[cli]»

# Download a specific quantization (e.g., the recommended Q5_1):
huggingface-cli download alekringtonnn-ai/zubr-whisper \
  whisper-small-q5_1.gguf

# Download the entire repository (all 6 quantizations, ~1 GB):
huggingface-cli download alekringtonnn-ai/zubr-whisper

By default, files are saved to ~/.cache/huggingface/hub/.

Method 2: Direct links (browser / wget / curl)

# The recommended version (Q5_1)
curl -L -o whisper-small-q5_1.gguf \
  «https://huggingface.co/alekringtonnn-ai/zubr-whisper/resolve/main/whisper-small-q5_1.gguf»

# Maximum quality (Q8_0)
curl -L -o whisper-small-q8_0.gguf \
  «https://huggingface.co/alekringtonnn-ai/zubr-whisper/resolve/main/whisper-small-q8_0.gguf»

The same links also work in a browser: open the file page and click «Download», or append ?download=true to the URL.

Method 3: Directly from whisper.cpp (with an HF base)

whisper.cpp can download GGUF models from Hugging Face using the —model flag:

whisper-cli —model zubr-whisper —language ru -f audio.wav

A more reliable path is to download the file via Method 1 or 2 and pass the path to it with the -m flag (see the «How to Run» section).

—

🚀 How to Run

Installing whisper.cpp

# macOS (Homebrew)
brew install whisper-cpp

# Windows
winget install whisper-cpp
# or download a prebuilt binary: https://github.com/ggml-org/whisper.cpp/releases

# Building from source (Linux)
git clone https://github.com/ggml-org/whisper.cpp
cd whisper.cpp
cmake -B build && cmake —build build —config Release

Preparing audio

whisper.cpp expects 16 kHz WAV. Convert with ffmpeg:

ffmpeg -i audio.mp3 -ar 16000 -ac 1 -c\:a pcm_s16le audio16k.wav

Transcription

# Basic launch:
whisper-cli -m whisper-small-q5_1.gguf -f audio16k.wav -l ru

# Useful flags:
whisper-cli \
  -m whisper-small-q5_1.gguf \
  -f audio16k.wav \
  -l ru \                       # audio language (ru = Russian)
  -otxt \                       # output to .txt (also: -osrt, -ovtt, -olrc, -oj)
  -of transcript                # result file name

The result appears as transcript.txt / transcript.srt / transcript.json (JSON includes per-segment timecodes).

Via Python (pywhispercpp)

pip install pywhispercpp

import pywhispercpp as whisper

model = whisper.Whisper(‘whisper-small-q5_1.gguf’)
segments = model.transcribe(‘audio16k.wav’, lang=‘ru’)

for segment in segments:
    print(f»[{segment.start:.1f}s → {segment.end:.1f}s] {segment.text}»)

—

⚙️ Recommended Parameters

Parameter	Value	Comment
-l (language)	ru / en / auto	Audio language; specifying it explicitly improves quality
—temperature	0.0–0.2	For transcription, the lower the more stable
-t (threads)	number of CPU cores	More threads = faster recognition
-otxt / -osrt / -ovtt / -oj	output format	Text, subtitles, JSON with timecodes

—