Skip to content
Audio

StreamVoice speech models

compact speech recognition models for mobile devices

2 min readupdated

A collection of lightweight automatic speech recognition (ASR) models built on the NVIDIA FastConformer architecture (115 million parameters) for the StreamVoice Android app. Optimized for the ggml format and quantized to 5 bits (q5_0), they run efficiently on mobile hardware while preserving punctuation and capitalization. The repository includes separate model files for multiple languages, including Russian, English, and several other European languages. This release is aimed at mobile developers and hobbyists who need fast, offline audio transcription.

Key features

  • Multilingual support: Includes separate files for Russian, Arabic, French, German, and other languages.
  • ggml optimization: Original NeMo model weights are converted to ggml format with q5_0 quantization for mobile chips.
  • Compact footprint: At 115M parameters, the models deliver high accuracy with minimal resource overhead.
  • Punctuation restoration: Automatically adds punctuation and capitalization to transcriptions.

What you can do

  • Integrate into Android apps: Load the model files into the StreamVoice mobile app or similar ggml-compatible ASR clients.
  • Run offline transcription: Transcribe Russian voice notes and audio messages locally without sending data to the cloud.
  • Benchmark performance: Test the inference speed of 5-bit quantized FastConformer models on mobile hardware.

Example

These models are designed for use inside the StreamVoice Android app rather than standard Hugging Face Python scripts. You work directly with prebuilt binary files using the .bin extension (for example, streamvoice-asr-ru-q5_0.bin for Russian).

Read moreDetails
• Base models: NVIDIA FastConformer Hybrid Large PC (CC-BY-4.0 license).
• Architecture: 115M parameters, converted from NeMo to ggml format using a TDT decoder.
• Quantization: 5-bit (q5_0); 4-bit quantization was skipped due to accuracy degradation.
• Limitations: No additional fine-tuning was performed; the files are tailored specifically for the app's mobile engine.
Discuss Telegram 0

Will it run?

Runs in Q5_0

Fits without quantisation.

  • Q5_0weights 0.1 GB · ready files0.6 GB
  • FP8 / INT8recommendedweights 0.0 GB0.5 GB
  • 4-bitweights 0.0 GB0.5 GB

Tip: go 8-bit.FP8 / INT8 is within about a percent of full precision in quality, takes half the memory and runs faster on RTX 40/50 and H100. Full precision only pays off for fine-tuning.

What to get for this model

Massed ComputeOpen Massed Compute

Your own GPU: FP8 / INT8 — from RTX 2060 · Q5_0 — from RTX 2060

Massed Compute hourly prices, November 2026. The discounted price applies with the coupon at checkout.

Estimate: weights plus working memory. Long videos, high resolutions and long contexts need more.

Run it on Massed Compute

The script downloads the model to the rented machine. How to run it is in the model card.

  1. 1. Rent a machine with a fitting GPU

    Open Massed Compute
  2. 2. Paste the script into the machine's terminal

    Connect over SSH (the Massed Compute panel shows the command) or open a terminal on the machine's desktop.

    HF_TOKEN / CIVITAI_TOKENThe script starts with token lines: paste yours between the quotes. The Hugging Face token is for gated models, the Civitai one for Civitai models.

    # ahmedmohamedra2002/streamvoice-speech-models
    mkdir -p ~/models
    curl -LsSf https://astral.sh/uv/install.sh | sh
    source "$HOME/.local/bin/env"
    uv tool install "huggingface_hub[cli]"
    hf download ahmedmohamedra2002/streamvoice-speech-models --local-dir ~/models/streamvoice-speech-models
    # The model is in ~/models. How to run it is in the model card (link above).
  3. 3. Use it

    The model is in ~/models. How to run it is in the model card (link above).

Stop the machine in the panel when you are done: it bills by the minute while it runs.

How to speed it up

Picked for this model type. Choose your hardware above to see what fits you exactly.

  • FlashAttention 44.0.0b40from November 18, 2026new release

    fastest exact attention on B200

    Kernels for Blackwell (beta). Exact attention without quantisation — fine for everything, LLMs included.

    How to enable: pip install flash-attn-4; vLLM and SGLang pick it up themselves.

    Earlier generations (2)
    • FlashAttention 3 — The Hopper (H100/H200) version with FP8.
    • FlashAttention 2 — The standard for RTX 30/40/50 and A100: exact attention without the memory cost of long contexts.
  • torch.compile2.15.0from October 28, 2026new release

    10–30% for free

    Compiles the model for your GPU. Slower first run, faster afterwards.

    How to enable: model = torch.compile(model); in ComfyUI the TorchCompileModel node.

Model family

New posts about versions of streamvoice-speech-models (quants, finetunes, LoRAs) will show up under the bell.

No known relatives of this model yet. All derivatives on Hugging Face

About the model

ahmedmohamedra2002/streamvoice-speech-models
Author
ahmedmohamedra2002
Category
Audio
Released
Posted

On NeyroLearn

Hugging Face

00

Ask AI about this model

Answered by Gemini 4 · the chat is kept only in this tab

Comments