instinct1912/instinct-stt-1

Instinct STT 1

Instinct STT 1 is a 585M-parameter multilingual CTC speech-recognition model for English, Russian, Uzbek, and Kazakh. Kyrgyz is retained as an evaluated zero-additional-training language. The model uses a shared automatic-language character decoder; callers do not need to provide a language tag.

The release contains the inference-only weights, configuration, and custom Transformers model code. It does not contain optimizer or scheduler state.

Usage

Install the runtime dependencies:

pip install "torch==2.10.*" "torchaudio==2.10.*" "transformers==5.*" hydra-core omegaconf soundfile

Request access on this model page, authenticate with Hugging Face, and run:

import os

from transformers import AutoModel

model = AutoModel.from_pretrained(
    "instinct-org/instinct-stt-1",
    trust_remote_code=True,
    token=os.environ.get("HF_TOKEN"),
)
model.eval()

text = model.transcribe("audio.wav")
print(text)

The included inference.py provides the same flow as a command-line program:

python inference.py audio.wav --device cuda

For a gated deployment, keep HF_TOKEN on the server. Never send it to a web or Telegram client.

Audio

The model operates at 16 kHz and accepts ordinary audio paths through model.transcribe. For production services, decode uploads to a supported audio file first and reject malformed or empty inputs. Long recordings should be segmented before transcription; this release is a full-context model, not a stateful streaming model.

Evaluation

Greedy-decoding WER on the held-out FLEURS evaluation used for the C1 release:

Language	WER (%)
English	9.1085
Russian	3.3296
Uzbek	7.4398
Kazakh	4.2836
Kyrgyz	5.9343

These scores are dataset-specific and should not be interpreted as guaranteed performance on arbitrary microphones, accents, domains, or noise conditions.

Release integrity

• Training checkpoint step: 42,690
• Full training checkpoint SHA-256: 5ab4f27b23e66002905831d112b66d0e1846fe91b5b2ffb266fecf50e52bb89f
• Inference tensor count: 824
• Inference weights SHA-256: 56c9001dc89bc85c5eb738f44b03c53fc7bef7a873b5c592dbfbcadfde12507d

The machine-readable export-receipt.json binds the published inference files to the retained training checkpoint.