Skip to content
Audio

lolo_Q8y_voice

The first TTS model for the Kuwaiti dialect

2 min readupdated

The lolo_Q8y_voice model by developer قُصَي الدُّعَيْج (Qusai Al-duaij) handles Text-to-Speech (Text-to-Speech) tasks with natural sound and broadcast-quality studio output in the authentic Kuwaiti dialect (24,000 Hz). The architecture relies on DiT Flow-Matching (base models F5TTS_v1_Base / Habibi-TTS) and the Vocos 24 kHz vocoder, with training taking place in two stages using around 67 hours of labeled audio data in total. A key feature of the model is strict phonetic marking of specific sounds (the Cairo-style jim گ and the Persian jim چ) alongside the preservation of standard Arabic sounds. The release is part of the lolo AI Family of models and will benefit voice bot developers, podcast creators, and researchers of regional Arabic dialects.

Key features

  • Studio quality: Generation of audio signals with a sampling rate of 24,000 Hz in Mono PCM format.
  • Two-stage training: Adaptation on 62 hours of general broadcast audio and an additional fine-tuning phase (Stage 2 Pure Kuwaiti Lock) on 4.78 hours of pure dialect speech.
  • Phonetic accuracy: Support for correct pronunciation and differentiation of the letters گ (G) and چ (Ch) in Kuwaiti vocabulary.
  • Built-in normalizer: The kuwaiti_normalizer.py utility for automatically converting standard Arabic text into phonetic orthography.

What you can do

  • Voiceover generation: Synthesize audio tracks for podcasts, dialogues, and radio scripts in the Kuwaiti dialect.
  • Bot integration: Connect the model weights via the F5-TTS library to build local voice assistants.
  • Text normalization: Apply the normalization script to prepare scripts accounting for local dialect features prior to generation.

Example

from huggingface_hub import hf_hub_download
from kuwaiti_normalizer import normalize_kuwaiti_tts_text

repo_id = "qalduaij/lolo_Q8y_voice"
ckpt_path = hf_hub_download(repo_id=repo_id, filename="lolo_Q8y_voice.safetensors")
vocab_path = hf_hub_download(repo_id=repo_id, filename="vocab.txt")

raw_text = "مساكم الله بالخير، شلونج وشخبارج؟ قاعد اقول لكم سالفة كويتية أصيلة باجر."
kuwaiti_text = normalize_kuwaiti_tts_text(raw_text)
print(kuwaiti_text)
Read moreDetails
Benchmarks: 62.01 hours (21,816 clips, 22,950 updates) in Stage 1 and 4.78 hours (1,867 clips, 1,416 updates) in Stage 2.
File size: The model weights lolo_Q8y_voice.safetensors take up 1.26 GB.
License: lolo-q8y-voice-inference-only-nc-nd-1.0 (inference only, no retraining rights, restructuring, or voice cloning without consent).
System requirements: The f5-tts library, huggingface_hub, and a reference audio file lasting 5–10 seconds.
Discuss Telegram 0 13

Will it run?

Runs in BF16

Fits without quantisation.

  • BF16weights 1.3 GB · ready files1.9 GB
  • FP8 / INT8recommendedweights 0.6 GB1.2 GB
  • 4-bitweights 0.4 GB0.9 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 · BF16 — 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.

    # qalduaij/lolo_Q8y_voice
    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 qalduaij/lolo_Q8y_voice --local-dir ~/models/lolo_Q8y_voice
    # The model is in ~/models. How to run it is in the model card (link above).
    Add more LoRAs to lolo_Q8y_voice
  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

  1. F5-TTSfamily root
  2. Habibi-TTS
  3. lolo_Q8y_voice

New posts about versions of F5-TTS (quants, finetunes, LoRAs) will show up under the bell.

Quantised versions 6All on Hugging Face
Fine-tunes 12All on Hugging Face
LoRAs and adapters 3All on Hugging Face

About the model

qalduaij/lolo_Q8y_voice
Author
qalduaij
Category
Audio
Released
Posted

On NeyroLearn

Hugging Face

130

Ask AI about this model

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

Comments