Skip to content
Audio

sava-telephony

fine-tune of the SAVA streaming model for telephone lines

2 min readupdated

The model is based on the SAVA architecture (25.5 million parameters, Mamba-2 and sliding-window attention) by Inworld and is optimized for low-quality audio over telephone lines. The original SAVA loses accuracy on telephony, so this fine-tune was trained on simulated telephone lines (8 kHz, codecs, and filtering). As a result, end-of-turn detection accuracy on a simulated landline increased from 89.5% to 93.0%, and on mobile (GSM) from 88.5% to 92.4%. The model is useful for developers of voice assistants, bots, and telephony systems who need fast real-time speech and pause detection.

Key features

  • Fine-tune for telephony: significantly better than the original at recognizing speech and pauses through telephone communication channels and codecs.
  • ONNX and quantization: PyTorch weights, standard ONNX, and an 8-bit quantized ONNX version of only 27 MB are available.
  • High speed: a single 40 ms frame is processed on one CPU thread of an Apple M4 Max in 1.2–1.5 ms.
  • Compatibility: fully retains the original streaming interface and architecture of the upstream model.

What you can do

  • Integration into a voice bot: embed the model into a call processing pipeline to quickly determine when the interlocutor has finished speaking.
  • Audio stream processing: connect inference via ONNX Runtime to handle real-time calls with upsampling to 16 kHz.
  • Hardware optimization: use the 8-bit quantized version sava-telephony-int8.onnx to reduce memory consumption on the server.

Example

import numpy as np
import onnxruntime as ort

session = ort.InferenceSession("sava-telephony-int8.onnx", providers=["CPUExecutionProvider"])
state = np.zeros((1, 798707), dtype=np.float32)

def step(frame: np.ndarray) -> tuple[float, float]:
    global state
    _, lookahead, turn, state = session.run(None, {"pcm": frame[None], "state": state})
    sigmoid = lambda x: 1.0 / (1.0 + np.exp(-x))
    return float(sigmoid(lookahead[0, 0])), float(sigmoid(turn[0, 0, 0]))
Read moreDetails
• Architecture: Mamba-2 and sliding-window attention, 25.5 million parameters.
• Accuracy (End-of-turn): wideband 93.5%, landline 93.0%, mobile 92.4%.
• Formats: safetensors (102 MB), ONNX (103 MB), INT8 ONNX (27 MB).
• Input data: 16 kHz mono float32 (8 kHz calls require prior upsampling).
• Limitations: improvement measured on simulated lines; testing on your own real calls is recommended.
Discuss Telegram 0

Will it run?

Runs in INT8

Fits without quantisation.

  • INT8recommendedweights 0.0 GB · ready files0.5 GB
  • 4-bitweights 0.0 GB0.5 GB

Tip: go 8-bit.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: INT8 — 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.

    # matthiastl/sava-telephony
    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 matthiastl/sava-telephony --local-dir ~/models/sava-telephony
    # The model is in ~/models. How to run it is in the model card (link above).
    Add more LoRAs to sava-telephony
  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 sava (quants, finetunes, LoRAs) will show up under the bell.

Quantised versions 2All on Hugging Face
Fine-tunes 1All on Hugging Face

About the model

matthiastl/sava-telephony
Author
matthiastl
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