Skip to content
Audio

AksharaVP-17M

ultra-lightweight 17.6M-parameter TTS with voice cloning

2 min readupdated

AksharaVP-17M by sourojitdotcom is a compact TTS (text-to-speech) model that synthesizes Telugu, Hindi, and Indian English at a 24 kHz sample rate. Under the hood, it combines an 8-layer transformer (15M parameters), a 4-step flow-matching latent frame generator, and Kyutai's Mimi decoder (2.6M parameters) — all packed into a single checkpoint with no external vocoders or phonemizers. Compared to larger alternatives like Hima TTS nano and Meta MMS-TTS, its main advantage is its tiny footprint (35.3 MB in f16 GGUF) paired with zero-shot voice cloning (copying a speaker's timbre from a few seconds of audio without fine-tuning). It is a practical fit for mobile, local, and browser-based apps running without a GPU.

Key features

  • Ultra-small footprint: just 17.6M parameters in the main checkpoint (plus 9.7M for the optional voice cloning encoder), making it one of the lightest multilingual TTS models available.
  • Fast CPU inference: generates speech 2.6–2.9x faster than real time on a single standard CPU thread, and up to 5.5x faster on four threads.
  • 12 voices and zero-shot cloning: ships with 12 built-in voices, and adding clone_encoder.safetensors lets you clone any voice from an audio clip of up to 5 seconds, including cross-lingual transfer.
  • Runs right in the browser: ported to ONNX Runtime Web (WebAssembly) to run entirely client-side without sending audio to a server.

What you can do

  • Embed offline voiceovers for notifications, articles, or navigation into mobile and desktop apps with tight memory budgets.
  • Deploy a client-side voice interface directly on a web page via WebAssembly with zero server GPU costs.
  • Personalize speech in bots or local assistants by cloning a target voice from a short, few-second reference clip.

Example

from aksharavp_tts import AksharaVP

tts = AksharaVP(".")
Read moreDetails
Benchmarks (FLEURS dev, 100 utterances): CER (character error rate) is 6.04% on Telugu, 4.57% on Hindi, and 4.37% on Indian English.
UTMOS naturalness score (1–5 scale): 2.64 (te), 2.73 (hi), 3.19 (en) — speech is intelligible, though flatter than larger TTS models.
Hardware and formats: requires no GPU; available in safetensors, f32 GGUF (70.6 MB), and nearly bit-identical f16 GGUF (35.3 MB).
Limitations: supports only the three listed languages (Telugu, Hindi, English) and trails heavier models in intonation naturalness.
Examples and videos
Discuss Telegram 0 34

Will it run?

Runs in GGUF F32

The most precise of the ready quantised files fits.

  • GGUF F32weights 0.1 GB · ready files0.6 GB
  • F32weights 0.1 GB · ready files0.6 GB
  • GGUF F16weights 0.0 GB · ready files0.5 GB
  • BF16weights 0.0 GB0.5 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 · GGUF F32 — 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.

    # sourojitdotcom/AksharaVP-17M
    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 sourojitdotcom/AksharaVP-17M --local-dir ~/models/AksharaVP-17M
    # The model is in ~/models. How to run it is in the model card (link above).
    Add more LoRAs to AksharaVP-17M
  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 AksharaVP-17M (quants, finetunes, LoRAs) will show up under the bell.

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

About the model

sourojitdotcom/AksharaVP-17M
Author
sourojitdotcom
Category
Audio
Released
Posted

On NeyroLearn

Hugging Face

340

Ask AI about this model

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

Comments