gigaam-multilingual-mlx 0.1.2

Native offline speech-to-text for Russian, Kazakh, Kyrgyz, and Uzbek on Apple Silicon with GigaAM and MLX

GigaAM-Multilingual MLX

English · Русский

Offline speech-to-text for Russian, Kazakh, Kyrgyz, and Uzbek on Apple Silicon. This is an independent native MLX port of GigaAM for Apple Silicon, based on the official GigaAM-Multilingual model. This project is not an official release of the GigaAM authors.

PyPI package · Model Collection · latest release

gigaam-multilingual-mlx runs local multilingual speech recognition without PyTorch, ONNX Runtime, Core ML, or a cloud ASR service. Model weights are downloaded separately from Hugging Face and reused from the standard Hub cache.

Install and transcribe in 60 seconds

Requirements: an Apple Silicon Mac, macOS 14 or newer, 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 the Python API inside a uv-managed project, use uv add gigaam-multilingual-mlx. A regular python -m pip install gigaam-multilingual-mlx remains supported. The canonical gigaam-multilingual-mlx transcribe ... command and the explicit gigaam-stt transcribe ... form remain available for compatibility.

The first run downloads the default INT8 model at the immutable v0.1.0 release tag. Later runs reuse the cached files.

Why GigaAM STT?

• Strong measured transcription quality for the four core GigaAM languages, especially Kazakh, Kyrgyz, and Uzbek.
• Native MLX inference on Apple Silicon without PyTorch, ONNX Runtime, Core ML, or a cloud ASR service.
• A 699 MB default model with substantially lower process memory than the tested MLX Whisper checkpoints.
• Fully local operation after the first model download, with a simple gigaam-stt AUDIO CLI and TXT, JSON, SRT, or VTT output.

Model selection: GigaAM MLX, MLX Whisper, and MLX Parakeet

Corpus WER uses pinned FLEURS test selections. Resource columns use one Russian five-minute WAV on a 14-inch MacBook Pro with Apple M4 Pro and 48 GB unified memory. 5-min WAV is the median of five runs after model load; Peak RAM is whole-process peak RSS; model size is the weight file on disk.

✓ column leader · ◇ Pareto frontier · ★ recommended default. Lower is better. English is an appendix. Parakeet does not officially support Kazakh, Kyrgyz, or Uzbek, so those cells are N/A.

◇ Pareto frontier — start here

The frontier is computed among MLX candidates using equal-weight macro WER over Russian, Kazakh, Kyrgyz, and Uzbek plus 5-min WAV time, Peak RAM, and model size. Original GigaAM is a PyTorch/MPS reference baseline; English is excluded as an appendix; models missing a core language are ineligible.

Pareto MLX variant	Core macro WER	5-min WAV	Peak RAM	Model size	Best fit
◇ GigaAM MLX FP16	5.066%	✓ 1.952s	1.350 GB	1.171 GB	fastest; reference-port fidelity
◇ ★ GigaAM MLX INT8	5.070%	2.036s	0.877 GB	0.699 GB	recommended quality / speed / footprint balance
◇ GigaAM MLX INT6	5.069%	2.195s	0.755 GB	0.573 GB	smaller footprint with near-INT8 quality
◇ GigaAM MLX INT4	5.219%	2.563s	✓ 0.626 GB	✓ 0.447 GB	minimum model size and Peak RAM

All compared models

Model / variant	RU WER	KZ WER	KY WER	UZ WER	EN WER	5-min WAV	Peak RAM	Model size
Original GigaAM PyTorch/MPS	✓ 2.995%	✓ 4.325%	✓ 5.553%	✓ 7.310%	9.717%	2.789s	5.059 GB	2.342 GB
◇ GigaAM MLX FP16	✓ 2.995%	4.342%	5.560%	7.367%	9.726%	✓ 1.952s	1.350 GB	1.171 GB
◇ ★ GigaAM MLX INT8	3.013%	4.351%	5.582%	7.334%	9.734%	2.036s	0.877 GB	0.699 GB
◇ GigaAM MLX INT6	3.013%	4.385%	5.568%	✓ 7.310%	9.848%	2.195s	0.755 GB	0.573 GB
◇ GigaAM MLX INT4	3.234%	4.377%	5.768%	7.497%	9.883%	2.563s	✓ 0.626 GB	✓ 0.447 GB
MLX Whisper large-v2	3.855%	39.468%	92.582%	95.551%	4.194%	14.287s	3.733 GB	3.083 GB
MLX Whisper large-v3	3.123%	32.778%	86.680%	87.981%	✓ 4.098%	18.214s	3.765 GB	3.084 GB
MLX Whisper large-v3-turbo	3.549%	20.468%	84.078%	108.931%	4.579%	6.722s	1.898 GB	1.614 GB
MLX Parakeet TDT 0.6B v3	4.961%	N/A	N/A	N/A	4.928%	3.843s	1.085 GB	2.508 GB

For Kazakh, Kyrgyz, and Uzbek, paired bootstrap 95% confidence intervals support lower WER for INT8 than all three Whisper references. Russian is close: the measured INT8 advantage is supported against large-v2 and v3 Turbo, but not against large-v3. INT8 also has 39.3% lower Russian WER than Parakeet in this suite (paired-bootstrap 95% CI: 32.7–45.6%). Whisper and Parakeet are better on the English appendix; Whisper supports many more languages and speech-to-English translation. Parakeet does not officially support Kazakh, Kyrgyz, or Uzbek, so those cells are N/A rather than out-of-scope measurements.

INT8 was 3.30× faster than Whisper v3 Turbo, 7.02× faster than large-v2, and 8.94× faster than large-v3 on this input, with 54–77% lower peak RSS. It was also 1.89× faster than Parakeet, with 19.2% lower peak RSS and 72.1% smaller weights. See the multilingual benchmark for all languages, FP16/INT6/INT4, CER, p95, confidence intervals, hashes, and limitations.

Model variants

Recommended default: INT8 g64. It preserves near-FP16 quality while giving the strongest measured balance of download size, process memory, and speed.

Variant	Role	Model size	Hugging Face repository
INT8 g64	recommended default, lowest-risk quantized	699 MB	ai-babai/gigaam-multilingual-mlx-int8-g64
FP16	reference MLX artifact	1.17 GB	ai-babai/gigaam-multilingual-mlx
INT6 g64	compact balanced	573 MB	ai-babai/gigaam-multilingual-mlx-int6-g64
INT4 g64	smallest opt-in	447 MB	ai-babai/gigaam-multilingual-mlx-int4-g64

All variants are grouped in the GigaAM Multilingual MLX Collection. The technical upstream branch name large_ctc remains artifact metadata; it is not part of the public project name.

Select another variant without knowing its repository name:

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

Inputs and outputs

ffmpeg provides decoding for WAV, FLAC, MP3, M4A, and audio tracks in common video containers. Inputs are converted to mono 16 kHz PCM before inference.

# Write next to an explicit path
gigaam-stt interview.mp3 --format srt --output interview.srt

# Generate OUTPUT_DIR/INPUT_STEM.vtt
gigaam-stt recording.mov --format vtt --output-dir transcripts/

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

Supported output formats are txt, json, srt, and vtt. CTC word timestamps are estimates derived from emission frames. They are monotonic and useful for navigation/subtitles, but they are not forced-alignment timestamps.

Long inputs are processed in deterministic overlapping chunks. The inference graph and tensors are reused between chunks instead of accumulating model state.

Cache, local models, and offline use

By default, Hugging Face stores model snapshots under ~/.cache/huggingface/hub. Use the normal Hub variables such as HF_HOME to move that cache, or pass an explicit directory:

gigaam-stt audio.wav --cache-dir /Volumes/ML/huggingface
gigaam-stt audio.wav --offline
gigaam-stt audio.wav --model /path/to/portable/model
gigaam-stt audio.wav \
  --model ai-babai/gigaam-multilingual-mlx-int8-g64 \
  --revision v0.1.0

--offline never contacts the Hub and fails clearly if the requested snapshot is not already cached. To remove only downloaded model data, use the Hugging Face CLI cache tools or delete the corresponding models--ai-babai--gigaam-* cache folders; do not remove unrelated Hub cache entries.

Python API

Existing local-directory loading remains supported, and Hub variants use the same function:

import mlx.core as mx
from gigaam_multilingual_mlx import load_model

model = load_model(variant="int8")  # downloads/reuses ai-babai/...-int8-g64@v0.1.0
audio = mx.zeros((1, 16_000), dtype=mx.float32)
log_probs, lengths = model(audio)
text = model.greedy_decode(log_probs, lengths)[0]["text"]

local_model = load_model("/path/to/portable/model")

Support matrix

Scope	Status
Apple Silicon supported by MLX, macOS 14+, native ARM Python	supported
M1–M5 generations	expected compatible; not benchmarked individually
Apple M4 Pro, 48 GB, macOS 15.7.7	release benchmark machine
Intel Mac, Linux, Windows, iOS	not supported in v0.1.0

The package reports an actionable error on unsupported platforms or when ffmpeg is missing.

Validation and benchmark

The accepted model artifacts use source revision 3905cd51c3ed4e88c8edf33f3302969ba480a327. Strict FP32 conversion reached 100% greedy-token agreement on the frozen parity set. Every production artifact reloads strictly with its recorded SHA-256.

The v0.1.0 release quality suite uses pinned subsets of FLEURS ru_ru, Russian LibriSpeech, and SOVA RuDevices with one normalization rule and identical greedy decoding. The release report publishes Original PyTorch/MPS, FP16, INT8, INT6, and INT4 side by side, including WER/CER, load/cold/warm timing, RTF, peak resident memory, MLX memory, and swap. It contains no private/user audio.

Results at a glance

Apple M4 Pro (48 GB), macOS 15.7.7. WER columns are FLEURS / Russian LibriSpeech / SOVA corpus WER. Runtime is the warm median for the same public 5-minute input; download time is excluded. Lower is better except speedup.

Implementation	Weights	WER: FLEURS / RuLibri / SOVA	Load	5-min warm	Speed vs Original	Peak RSS
Original PyTorch/MPS	2.342 GB	6.271% / 5.911% / 12.448%	6.066s	3.023s	1.00×	5.019 GB
MLX FP16	1.171 GB	6.271% / 5.929% / 12.431%	0.947s	2.089s	1.45×	1.348 GB
MLX INT8 g64 (default)	0.699 GB	6.292% / 5.911% / 12.481%	0.578s	2.332s	1.30×	0.879 GB
MLX INT6 g64	0.573 GB	6.249% / 6.022% / 12.465%	0.478s	3.142s	0.96×	0.755 GB
MLX INT4 g64	0.447 GB	6.358% / 6.003% / 12.581%	0.378s	2.750s	1.10×	0.628 GB

INT8 is the default for its measured quality, size, and memory balance. FP16 is the fastest measured variant on the representative 5-minute input; quantization does not guarantee a speedup for every duration. Peak RSS is whole-process resident memory, not model-only or additional GPU memory.

See benchmarks/README.md for the reproducible public-only methodology and docs/benchmark-v0.1.0.md for release results.

Limitations

• The v0.1.0 release exposes greedy CTC decoding; it does not add beam search.
• Timestamps are approximate CTC emission times.
• Training, fine-tuning, diarization, microphone streaming, and a local HTTP server are outside this release.
• Quality is strongest for Russian, Kazakh, Kyrgyz, and Uzbek; upstream reports moderate English quality. See the upstream model card for the authoritative scope.
• Model and dataset biases from the upstream training/evaluation material remain.

Developer tools

Conversion, quantization, parity, quality evaluation, and benchmark commands are kept out of the production CLI and use optional dependencies:

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

Weights, datasets, generated audio, and raw benchmark outputs must stay outside the Git repository.

License, attribution, and citation

The code in this repository is MIT-licensed. GigaAM code/model provenance and dataset notices are recorded in THIRD_PARTY_NOTICES.md. When using the model, cite the original GigaAM-Multilingual work from the upstream model card; project citation metadata is available in CITATION.cff.

Security reports and contributions are described in SECURITY.md and CONTRIBUTING.md.