Pocket TTS Indonesian

Pocket TTS Indonesian

Speaks Indonesian text aloud and can clone a voice from a short recording, running on a CPU with the 6-layer model. Made for developers and creators who need Indonesian speech and have the right to use the voice.

Made by anak10thn on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: Pocket TTS Indonesian turns Indonesian text into speech and can copy a voice from a short recording. It is based on kyutai's pocket-tts. Training used 502 hours of Indonesian speech from the LEMAS dataset. The repository holds a 6-layer model and a 24-layer one, and the card says to use the 6-layer model.

Who it suits: It suits developers and creators who need Indonesian voice-overs or a small speech model that runs without a GPU.

Industry | Who uses it | How they use it | Example
Software | App developers | Add Indonesian speech to an app | Read a greeting aloud in Indonesian
Media | Video and podcast makers | Produce Indonesian voice-overs | Narrate a script in a voice from a short recording
Education | Language teachers | Make spoken practice material | Generate audio for short Indonesian sentences
Research | Speech researchers | Study small distilled speech models | Compare the 6-layer and 24-layer models

If you need another language, or you do not have permission to use the voice you want to copy, you can skip it.

Where to use it: It runs from the command line or from Python, and the 6-layer model works on a CPU.

• The pocket-tts command line, run through uvx
• Python scripts using the TTSModel class
• Windows, macOS or Linux machines, with no GPU needed for the 6-layer model

Use case | Who | What they get back
Indonesian speech | Developers | A WAV file of the text spoken aloud
Voice cloning | Creators | Speech in a voice close to your recording
CPU-only speech | Teams without GPUs | Output at about 2.23 times real time

How to set it up: You need Python and the uvx tool to run the command, or the pocket_tts package for Python use. You also need a voice recording, or you can use one of the four sample voices in the repository.

1. Install the Hugging Face hub tools.
2. Download the model files if you want them locally.
3. Run the generate command, which fetches the config by its path.
4. Pass your own recording, or leave the voice out to use the default English voice.

pip install -U huggingface_hub
hf download anak10thn/pocket-tts-indonesian

How to use it: Run the generate command with a voice file, your text, the end-of-speech setting and an output path. The card says to pass the eos threshold of -6.0.

uvx pocket-tts generate \
    --config hf://anak10thn/pocket-tts-indonesian/indonesian_6l.yaml@17257664e384561c957b02ac92edd1a24807f0e5 \
    --voice your_voice.wav \
    --text "Selamat pagi, semoga hari Anda menyenangkan." \
    --eos-threshold -6.0 \
    --output-path out.wav

From Python, load the model, build a state from your recording, and generate audio.

from pocket_tts import TTSModel

model = TTSModel.load_model(
    config="hf://anak10thn/pocket-tts-indonesian/indonesian_6l.yaml@17257664e384561c957b02ac92edd1a24807f0e5"
)
state = model.get_state_for_audio_prompt("your_voice.wav")
audio = model.generate_audio(state, "Selamat pagi, semoga hari Anda menyenangkan.")

Tool | What it does
uvx pocket-tts generate | Makes a speech file from text and a voice prompt
--eos-threshold -6.0 | Sets the end-of-speech value the card recommends
TTSModel | Loads the model and generates audio in Python
hf download | Fetches the model files to your computer

Example: You pass your_voice.wav and the greeting "Selamat pagi, semoga hari Anda menyenangkan." to the generate command. You get out.wav, an Indonesian spoken version of that greeting in a voice like your recording. Only use voices you have the right to use.

For newcomers and experts: Voice cloning makes the model copy a speaker's voice from a sample. Start with one of the four sample voices and a short sentence. Write numbers as words, because digits are not in the vocabulary.

The small student model was distilled from the larger teacher with guidance baked in, so it gets guided quality in one pass. A quiet 48 kHz recording clones more distinctly than a louder compressed one. The tokenizer ignores letter case, and hyphens now work.

Prepare a voice recording
Write Indonesian text, numbers as words
Run the generate command
The 6-layer model speaks the text
Save the audio as a WAV file

Good to know: The card measured the student against the teacher on 153 cross-sentence pairs from 116 speakers, and the student scored a lower word error rate.

Title: Median word error rate on the eval set, lower is better
6-layer student: 9.09 percent
24-layer teacher: 16.67 percent

• Size: the 6-layer model is about 438 MB, against 1.27 GB for the teacher.
• Speed: about 2.23 times real time on a CPU, while the teacher runs at 0.69 times.
• Quality: voice similarity and clone distinctness depend a lot on the recording quality.
• License: CC BY 4.0.
• Cost to run: none named.
• Data it reaches: your voice recording and text, and the model files fetched from Hugging Face.
• Not for: languages other than Indonesian, or cloning voices without permission.

On Hugging Face, 13 November 2026

• Task: text to speech
• Library: pocket-tts
• Formats: Safetensors
• Downloads in the last 30 days: 0
• Likes: 10
• License: cc-by-4.0
• Languages: id
• Based on: kyutai/pocket-tts
• Quantized versions: 1
• Tags: pocket-tts, onnx, safetensors, text-to-speech, indonesian, bahasa-indonesia, voice-cloning

Links

• Model: https://huggingface.co/anak10thn/pocket-tts-indonesian
• Files: https://huggingface.co/anak10thn/pocket-tts-indonesian/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download anak10thn/pocket-tts-indonesian

Then follow the usage in its model card. Nothing in this kit runs by itself.