donglao-tts

Reference-voice text-to-speech with a simple Python API.

Research/pre-1.0 release. APIs may change between versions. Use reference voices only with the speaker's permission.

Recommended Python API

uv add donglao-tts

# Or with pip
python -m pip install donglao-tts

from donglao_tts import DongLaoTTS

tts = DongLaoTTS.from_pretrained("DongLao/DongLao-TTS")

waveform = tts.generate(
    "Xin chào mọi người, đây là Đông Lào TTS.",
    reference_audio="reference.wav",
    reference_text="Exact transcript of the speech in reference.wav.",
    output_path="output.wav",
)

print(tts.revision)
print(tts.sample_rate)
print(tuple(waveform.shape))

The first call downloads the model, tokenizer, and bundled MOSS codec. CUDA is selected when available; otherwise the model runs on CPU. Reuse the loaded tts object for subsequent calls.

For reproducible use, pin a tested model commit:

tts = DongLaoTTS.from_pretrained(
    "DongLao/DongLao-TTS",
    revision="6ba3003ccb8d938c2a725a4117084492909c9419",
    device="cuda",
)

Generation controls are available when needed:

waveform = tts.generate(
    "Text to synthesize.",
    reference_audio="reference.wav",
    reference_text="Exact reference transcript.",
    output_path="output.wav",  # optional
    max_frames=200,
    temperature=0.8,
    top_k=10,
)

After G2P conversion, the target is split on periods, synthesized sentence by sentence, and the resulting audio is concatenated in order. The returned waveform is a CPU PyTorch tensor with shape [channels, samples].

Multiple targets sharing the same reference can reuse its preprocessing with generate_batch():

waveforms = tts.generate_batch(
    ["First target.", "Second target."],
    reference_audio="reference.wav",
    reference_text="Exact reference transcript.",
)

generate_stream() yields audio decoded from consecutive RVQ-token groups:

for audio_chunk in tts.generate_stream(
    "First sentence. Second sentence.",
    reference_audio="reference.wav",
    reference_text="Exact reference transcript.",
    chunk_frames=5,
):
    consume(audio_chunk)

The AR KV-cache is preserved between groups. NAR completes the remaining RVQ layers and MOSS decodes each group while AR generation continues. At a 25 Hz codec rate, five frames represent approximately 200 ms of audio.

Runtime requirements

• Python >=3.10,<3.11
• Linux x86-64
• PyTorch and TorchAudio >=2.8.0,<3
• CUDA-capable GPU recommended; CPU inference is supported

Reference audio

Use a clean, consented recording containing one speaker. reference_text must match the spoken content exactly. Avoid music, overlapping speakers, clipping, long silence, and heavy reverb.

Verify the installation

donglao-smoke-test-hub \
  --repo-id DongLao/DongLao-TTS \
  --device cpu

For an end-to-end test, also pass --ref-audio, --ref-text, --target-text, and --output.

Responsible use

Obtain consent before using a voice, disclose synthetic audio when appropriate, and protect reference recordings and transcripts. Do not use the package for impersonation, fraud, harassment, or bypassing voice authentication.

Links

• Source and full documentation
• Published model
• Issue tracker
• Apache License 2.0