model_device

Model & Device Options

These arguments control which AI model is used and how it runs on your hardware.

Arguments

Flag	Description
--model_source	AI model backend (whisper, fasterwhisper, openvino).
--ram	Model size (choices: 1gb, 2gb, 3gb, 6gb, 7gb, 11gb-v2, 11gb-v3).
--ramforce	Force the script to use the selected RAM/VRAM model.
--fp16	Enable FP16 mode for faster inference (may reduce accuracy slightly).
--compute_type	Quantization of model while loading (default, int8, float16, etc.).
--device	Select device for inference (auto, cpu, cuda, intel-igpu, intel-dgpu, intel-npu).
--cuda_device	Select CUDA device index (default: 0).
--model_dir	Directory to store/download models.
--intelligent_mode	Automatically switch to larger model if accuracy is below threshold.

Details & Examples

--model_source

Choose AI model backend for transcription:

• whisper: OpenAI's original implementation (slowest, most compatible)
• fasterwhisper: Optimized C++ implementation (fastest, recommended)
• openvino: Intel optimization for Intel hardware (good for Intel CPUs/GPUs)

Example:

python synthalingua.py --model_source fasterwhisper

--ram & --ramforce

Choose a model size that fits your hardware. For example:

python synthalingua.py --ram 6gb

Model sizes and descriptions:

• 1gb: tiny (fastest, least accurate)
• 2gb: base (good balance)
• 3gb: small
• 6gb: medium
• 7gb: turbo (fast large model)
• 11gb-v2/11gb-v3: large (most accurate, slowest)

Use --ramforce to override automatic checks (use with caution).

--fp16

Enables half-precision mode for faster processing on supported GPUs.

--compute_type

Controls quantization of model while loading. Options include:

• default: Automatic selection
• int8, int8_float32, int8_float16, int8_bfloat16: 8-bit integer variants
• int16: 16-bit integer
• float16, bfloat16: 16-bit floating point variants
• float32: 32-bit floating point

--device & --cuda_device

Selects processing device for AI model inference:

• auto: Automatically selects best available device
• cuda: NVIDIA GPU (fastest)
• cpu: Processor (universally compatible)
• intel-igpu: Intel integrated graphics
• intel-dgpu: Intel discrete GPU
• intel-npu: Intel Neural Processing Unit

Example:

python synthalingua.py --device cuda --cuda_device 1

--model_dir

Change where models are stored/downloaded. Example:

python synthalingua.py --model_dir "C:/models"

--intelligent_mode

If enabled, the system will automatically determine if the current output is below accuracy threshold and switch to a larger model for improved transcription quality.

---

Back to Index