Models

Default model

Speech is recognized by GigaAM-v3 E2E RNN-T (gigaam-v3-e2e-rnnt) - an open Russian speech recognition model by SberDevices under the MIT license. It writes text with punctuation, capital letters and numbers as digits right away. The model runs through onnx-asr and ONNX Runtime, without PyTorch, with int8 weights by default.

Other models

Any onnx-asr model can be chosen by name with -m or the model argument of Transcriber; it is downloaded from Hugging Face on the first run. Models that recognize Russian:

Model	Features
gigaam-v3-e2e-rnnt, gigaam-v3-e2e-ctc	Punctuation, capital letters and numbers as digits; CTC is usually faster than RNN-T
gigaam-v3-rnnt, gigaam-v3-ctc	Lowercase without punctuation, numbers as words
gigaam-v2-rnnt, gigaam-v2-ctc	The previous GigaAM version
t-tech/t-one	T-Bank model for telephone speech
nemo-fastconformer-ru-rnnt, nemo-fastconformer-ru-ctc	NVIDIA NeMo FastConformer for Russian with punctuation
alphacep/vosk-model-ru, alphacep/vosk-model-small-ru	Vosk models; small is for low-end computers
onnx-community/whisper-large-v3-turbo	OpenAI Whisper multilingual model; much heavier than the others
nemo-parakeet-tdt-0.6b-v3	NVIDIA multilingual model

The full list and the model file formats are in the onnx-asr documentation. Which model is best for your recordings is shown by a quality evaluation on your own references:

stt-ru evaluate eval.csv -m gigaam-v3-e2e-rnnt -m t-tech/t-one -m alphacep/vosk-model-ru

Parameters

The model, device and speech detection parameters are shared by the transcribe, evaluate and serve commands.

Parameter	Default	Description
-m, --model	gigaam-v3-e2e-rnnt	onnx-asr model; can be repeated for evaluate
--quantization	int8	Weight quantization: int8 or none - original weights
--model-dir	-	Folder with model files; for evaluate only with one model
--device	auto	Device: auto, cpu, cuda, coreml, directml (see Installation)
--mono	off	Mix channels into one
--ffmpeg	-	Path to the ffmpeg program
--vad-threshold, --min-speech-ms, --max-speech-s, --min-silence-ms, --speech-pad-ms	see below	Speech detection parameters

Quantization

int8 weights are about four times smaller than the original ones and usually faster on a CPU. --quantization none takes the original weights; a quality evaluation with both values shows the difference on your recordings. If a model has no int8 weights, the original ones are loaded with a warning.

Model files on disk

--model-dir (the model_dir argument) takes the recognition model files from a folder: for example, a model downloaded by hand. If the folder does not exist, the model is downloaded into it. Silero VAD is still taken from the Hugging Face cache, so on a computer without internet it is simpler to copy the whole cache and point HF_HOME to it (see Installation).

stt-ru transcribe call.wav -m gigaam-v3-e2e-rnnt --model-dir /models/gigaam-v3-e2e-rnnt

Speech detection

stt_ru.vad.VadOptions

Before recognition the recording is cut into speech segments: the models accept segments of at most 20-30 s. Speech is detected by Silero VAD, and the defaults are those of Silero VAD and onnx-asr.

Command parameter	VadOptions field	Default	Description
--vad-threshold	threshold	0.5	Speech probability that starts a segment, from 0 to 1
-	neg_threshold	threshold - 0.15	Probability below which a segment ends
--min-speech-ms	min_speech_duration_ms	250	Shortest segment, ms: shorter ones are dropped
--max-speech-s	max_speech_duration_s	20	Longest segment, s: longer ones are cut
--min-silence-ms	min_silence_duration_ms	100	Shortest pause that separates segments, ms
--speech-pad-ms	speech_pad_ms	30	Padding at the segment edges, ms

• A noisy recording gives segments of noise - raise --vad-threshold, for example to 0.6.
• The model misses quiet speech - lower the threshold to 0.35-0.4.
• Utterances are split into segments that are too short - increase --min-silence-ms, for example to 300: segments become longer and subtitles less frequent.
• max_speech_duration_s must be at least one second longer than two speech_pad_ms.

>>> from stt_ru import Transcriber
>>> from stt_ru.vad import VadOptions

>>> transcriber = Transcriber(vad=VadOptions(threshold=0.6, min_silence_duration_ms=300))