GigaAM-Multilingual MLX

English · Русский

Fast, offline speech-to-text for Russian, Kazakh, Kyrgyz, and Uzbek on Apple Silicon. This is an independent native MLX port of GigaAM, based on the official GigaAM-Multilingual model.

PyPI · Hugging Face models · Latest release · Full benchmark

The chart compares WER, five-minute transcription time, peak process memory, and model size. Filled pills mark column leaders, outlined pills mark runners-up, ◇ marks the Pareto frontier, and ★ the recommended default. Lower is better.

Quick start

Requires an Apple Silicon Mac, macOS 14+, Python 3.12 or 3.13, uv, and ffmpeg.

brew install uv ffmpeg
uv tool install gigaam-multilingual-mlx
gigaam-stt meeting.m4a --output transcript.txt

For a uv-managed Python project:

uv add gigaam-multilingual-mlx

pip install gigaam-multilingual-mlx and the long gigaam-multilingual-mlx transcribe ... command remain supported.

Local transcription server

Use GigaAM MLX with applications that support the OpenAI transcription API:

uv tool install 'gigaam-multilingual-mlx[server]'
gigaam-stt serve

The server is now available at http://127.0.0.1:8000/v1:

curl http://127.0.0.1:8000/v1/audio/transcriptions \
  -F model=whisper-1 \
  -F file=@meeting.m4a

The same endpoint works with the OpenAI Python client (uv add openai):

from pathlib import Path
from openai import OpenAI

client = OpenAI(base_url="http://127.0.0.1:8000/v1", api_key="local")
with Path("meeting.m4a").open("rb") as audio:
    result = client.audio.transcriptions.create(model="whisper-1", file=audio)
print(result.text)

whisper-1 is a compatibility alias; the server still runs GigaAM Multilingual MLX. It is local-only by default. See the server guide for formats, model IDs, network access, configuration, and troubleshooting.

Why this port?

• Native MLX inference without PyTorch, ONNX Runtime, Core ML, or a cloud API.
• Strong measured quality for the four core GigaAM languages, especially Kazakh, Kyrgyz, and Uzbek.
• A 699 MB default model that used 0.877 GB peak process memory in the published five-minute benchmark.
• Local WAV, FLAC, MP3, M4A, and video transcription to TXT, JSON, SRT, or VTT.

On the tested Russian five-minute input, INT8 was 3.30× faster than Whisper v3 Turbo, 7.02× faster than Whisper large-v2, and 8.94× faster than Whisper large-v3. Whisper and Parakeet were better on the English appendix. These are single-machine public-corpus results, not a universal ASR leaderboard; see the full report for confidence intervals, commands, hashes, and limitations.

Model variants

Choose a variant without remembering its repository name:

gigaam-stt speech.wav --variant fp16
gigaam-stt speech.wav --variant int6 --format json
gigaam-stt models

The runtime downloads immutable model revision v0.1.0 on first use and then reuses the standard Hugging Face cache.

Common commands

# Subtitles
gigaam-stt interview.mp3 --format srt --output interview.srt

# Output directory; creates transcripts/recording.vtt
gigaam-stt recording.mov --format vtt --output-dir transcripts/

# Machine-readable text, timestamps, revision, and metrics
gigaam-stt sample.flac --format json --output sample.json

# Reuse a downloaded snapshot without network access
gigaam-stt audio.wav --offline

# Move the Hugging Face cache
gigaam-stt audio.wav --cache-dir /Volumes/ML/huggingface

Long inputs are processed in deterministic overlapping chunks. CTC word timestamps are approximate emission times, suitable for navigation and subtitles but not forced alignment.

Python API

import mlx.core as mx
from gigaam_multilingual_mlx import load_model

model = load_model(variant="int8")
audio = mx.zeros((1, 16_000), dtype=mx.float32)
log_probs, lengths = model(audio)
text = model.greedy_decode(log_probs, lengths)[0]["text"]

Local portable model directories are also supported: load_model("/path/to/model").

Compatibility and limitations

• Supported: Apple Silicon, native ARM Python, macOS 14+.
• M1–M5 are expected to be runtime-compatible; published performance numbers come from one 14-inch MacBook Pro with Apple M4 Pro and 48 GB memory.
• Not supported: Intel Mac, Linux, Windows, or iOS.
• This release uses greedy CTC decoding. It does not provide diarization, training, microphone streaming, or realtime HTTP/WebSocket streaming.
• Accuracy can degrade with noise, far-field speech, overlapping speakers, code-switching, or domains unlike the public evaluation data.

Reproducibility and development

The repository contains pinned public FLEURS manifests, compact benchmark results, and commands for GigaAM, MLX Whisper, and MLX Parakeet. No private audio, model weights, datasets, caches, or raw large benchmark outputs are stored in Git.

Developer-only conversion and evaluation tools use optional dependencies:

python -m pip install 'gigaam-multilingual-mlx[convert,quality]'
python -m gigaam_multilingual_mlx.dev_cli --help

Start with benchmarks/multilingual-v1/README.md for the protocol and docs/benchmark-multilingual-v1.md for the results.

Key numbers in text

This compact table is the text alternative to the GigaAM part of the image. Core macro WER averages Russian, Kazakh, Kyrgyz, and Uzbek; lower is better.

License and attribution

The MLX port is MIT-licensed. Upstream and dataset provenance is recorded in THIRD_PARTY_NOTICES.md. Please cite the original GigaAM-Multilingual work from the upstream model card and this project using CITATION.cff.

This project is not an official release of the GigaAM authors. Security reports and contributions are covered by SECURITY.md and CONTRIBUTING.md.