How to Run Kova TTS 1 Locally: a 1.25B-Parameter Voice Cloning Model for Narration

Kova TTS 1 is an open-weight, 1.25 billion parameter English text-to-speech model built on Llama 3.2 1B. It clones a voice from reference audio and a matching transcript, streams audio in under 200 milliseconds, and outputs 48 kHz mono audio. The license covers research and non-commercial use only.

Written by Priya Raghunathan, local-AI hardware reviewer. I check parameter counts, licenses and hardware requirements against the primary source before recommending an install path.

A new top-20 entrant on Hugging Face's trending list

Kova TTS 1 entered the top 20 of Hugging Face's trending text-to-speech list. At time of writing it has 9 likes and 889 downloads on the kova-ai/kova-tts-1 model page, released September 23, 2026, under a research and non-commercial license. The GitHub repo has 6 stars. The top linked Space is kova-ai/kova-tts-1 (4 likes), a browser demo of the model.

Specs at a glance

What it actually does

Kova TTS 1 takes a short reference clip of a voice, the exact transcript of what that clip says, and a new line of text, then speaks the new text back in the cloned voice without any fine-tuning step. The model grew out of Kova's own audiobook app, where the team needed to synthesize long-form narration for about 10,000 readers and wanted a voice that stayed pleasant to listen to for hours. That history shows up in the design: a 48 kHz audio decoder for cleaner playback than most open TTS models ship, a separate 13.7 million parameter aligner that produces word-level timestamps for subtitles or read-along highlighting, and automatic text normalization so numbers and abbreviations do not need to be spelled out by hand.

Install it

Kova TTS 1 ships as a GitHub repo managed with uv, not a plain pip package. You need Python 3.10 or newer and uv installed first.

That last command pulls the model weights, codec and aligner checkpoints and caches them locally. To try the model immediately without writing code, run the bundled browser demo instead. Add the demo and normalize extras so text normalization and the web interface are both available.

Generate speech from the command line

The CLI takes the text and writes a wav file. No reference voice means it falls back to the model's own default voice.

Clone a voice in Python

Voice cloning needs a reference clip and the exact transcript of what is said in it, the model uses the transcript to align its understanding of the reference audio. Reference audio can be 16 kHz or 32 kHz; Kova TTS 1 upsamples it to the 48 kHz it outputs.

from kova_tts import KovaTTS

tts = KovaTTS.from_pretrained()

voice = tts.clone(
    "reference.wav",
    transcript="Exactly what the reference clip says.",
)

wav = tts.generate("A new sentence in the reference voice.", voice)
tts.save(wav, "cloned.wav")

For streaming output instead of waiting on the full clip, Kova TTS 1 exposes the same generation over a local server with both WebSocket and HTTP Server-Sent Events endpoints, which is what gets the sub-200-millisecond numbers in the specs table above. Start the server with the CLI's serve command and connect to it from any client that speaks WebSocket or SSE.

Hardware and pretrained voices

Kova TTS 1 runs on NVIDIA CUDA, on Apple Silicon through a converted MLX checkpoint, and on CPU, though the model card describes CPU as slow for a 1.25 billion parameter model. If you do not have reference audio to clone, a separate repository holds five optional LoRA voice adapters you can load instead of cloning your own.

Known limitations

Kova TTS 1 speaks English only; there is no multilingual mode to fall back on. Voice cloning requires an exact transcript of the reference clip, so a rough or approximate transcript will degrade the clone. The bundled server is a development tool, not a production deployment, as Kova's own card points out. And because the license is research and non-commercial, you cannot deploy this in a paid product without getting a separate commercial license from Kova AI first.

Kova TTS 1 vs Sopro V2 Turbo

Both are voice-cloning text-to-speech models with a streaming mode, but they target different jobs. Sopro V2 Turbo is a 120 million parameter model licensed under Apache 2.0, runs faster than real time on a laptop CPU, and speaks English, Portuguese, French and German. Kova TTS 1 is over ten times larger at 1.25 billion parameters, speaks English only, and is licensed for research and non-commercial use unless you buy a license from Kova AI. In exchange, Kova TTS 1 outputs a higher 48 kHz sample rate and adds word-level timestamps out of the box, which fits its original job of long-form audiobook narration. Reach for Sopro V2 Turbo when you need a permissive license and multiple languages; reach for Kova TTS 1 when narration quality and timestamps matter more and the non-commercial license is acceptable.

FAQ

What is Kova TTS 1?

Kova TTS 1 is an open-weight, 1.25 billion parameter English text-to-speech model built on Llama 3.2 1B. It clones a voice from reference audio and a matching transcript, and outputs 48 kHz mono audio.

Can I use Kova TTS 1 commercially?

Not without a separate agreement. It ships under a research and non-commercial license. Commercial use requires a written license from Kova AI.

Do I need a GPU to run it?

No, it also runs on CPU and on Apple Silicon through MLX, but the model card describes CPU inference as slow at this parameter count. The published latency numbers were measured on an RTX 5090 and an H100.

What do I need to clone a voice?

A reference audio clip at 16 kHz or 32 kHz and the exact transcript of what is said in that clip. No fine-tuning step is required.

Where to go from here

• Want a permissively licensed, multilingual alternative? Run Sopro V2 Turbo locally covers a smaller voice-cloning model under Apache 2.0.
• Need fixed stock voices instead of cloning? Run Kokoro-82M locally covers a CPU-friendly TTS model with 54 voices across 8 languages.
• Browse other speech and audio models: the local AI model directory lets you compare specs side by side.