nur-dev/ait-piper-multilingual

AIT Piper Multilingual

Fast, offline multilingual text-to-speech for Kazakh, Russian, and English. The model runs through Piper/ONNX, contains 13 fixed voices, and supports incremental audio generation for real-time applications.

Model Characteristics

Property	Value
Languages	Kazakh (kk), Russian (ru), English (en)
Architecture	Piper/VITS
Runtime	ONNX Runtime
Sample rate	22,050 Hz
Audio	Mono PCM
Speakers	13 fixed voices
Voice cloning	Not supported
Streaming	Supported
Model size	Approximately 74 MB
Phonemizers	eSpeak NG: kk, ru, en-us

The bundled configuration uses the recommended moderately slower and more stable synthesis preset:

length_scale = 1.12
noise_scale = 0.55
noise_w_scale = 0.45

Voices

ID	Language	Voice
0	Kazakh	kk_F1
1	Kazakh	kk_F2
2	Kazakh	kk_F3
3	Kazakh	kk_M2
4	Kazakh	kk_emo_1263201035
5	Kazakh	kk_emo_399172782
6	Kazakh	kk_emo_805570882
7	Russian	ru_ruls_13587
8	Russian	ru_ruls_295
9	Russian	ru_ruls_8086
10	Russian	ru_ruls_8169
11	Russian	ru_ruls_9014
12	English	ljspeech_F1

For Russian production use, start with ru_ruls_8086 or ru_ruls_8169.

Download And Install

pip install "piper-tts>=1.4,<2" onnxruntime huggingface_hub

huggingface-cli download nur-dev/ait-piper-multilingual \
  --local-dir ./ait-piper-multilingual

Authentication is required because this repository is private. Keep the ONNX and .onnx.json files together with the same base filename.

Command-Line Inference

Kazakh:

echo "Бүгін Алматыда күн ашық, ауа райы жылы болады." | piper \
  --model ./ait-piper-multilingual/ait-piper-multilingual-medium.onnx \
  --language kk --speaker 3 --output-file kk.wav

Russian:

echo "Сегодня в Алматы солнечно и тепло." | piper \
  --model ./ait-piper-multilingual/ait-piper-multilingual-medium.onnx \
  --language ru --speaker 9 --output-file ru.wav

English:

echo "Real-time speech synthesis is ready." | piper \
  --model ./ait-piper-multilingual/ait-piper-multilingual-medium.onnx \
  --language en --speaker 12 --output-file en.wav

Always pass the correct language. The language selects the phonemizer; the speaker ID selects the voice.

Python Inference

import wave

from piper import PiperVoice, SynthesisConfig

model = "./ait-piper-multilingual/ait-piper-multilingual-medium.onnx"
voice = PiperVoice.load(model)

config = SynthesisConfig(
    speaker_id=3,
    language="kk",
    length_scale=1.12,
    noise_scale=0.55,
    noise_w_scale=0.45,
)

with wave.open("output.wav", "wb") as wav_file:
    voice.synthesize_wav(
        "Қазақ тіліндегі дыбыстау жүйесі жұмыс істеп тұр.",
        wav_file,
        syn_config=config,
    )

Load PiperVoice once at application startup and reuse it for requests.

Streaming Inference

synthesize yields signed 16-bit mono PCM chunks at 22,050 Hz:

from piper import PiperVoice, SynthesisConfig

voice = PiperVoice.load(
    "./ait-piper-multilingual/ait-piper-multilingual-medium.onnx"
)
config = SynthesisConfig(speaker_id=9, language="ru")

for chunk in voice.synthesize(
    "Потоковый синтез речи готов к работе.",
    syn_config=config,
):
    send_audio(
        chunk.audio_int16_bytes,
        sample_rate=chunk.sample_rate,
        sample_width=chunk.sample_width,
        channels=chunk.sample_channels,
    )

Chunk boundaries follow sentence boundaries. Split long input into complete sentences for lower first-audio latency.

HTTP Server

The repository includes serve.py, a minimal FastAPI service with WAV and streaming PCM endpoints.

cd ait-piper-multilingual
pip install -r requirements.txt
uvicorn serve:app --host 0.0.0.0 --port 8000 --workers 1

Generate a WAV file:

curl -s http://localhost:8000/v1/audio/speech \
  -H 'Content-Type: application/json' \
  -d '{
    "text": "Бүгін Алматыда ауа райы жылы.",
    "language": "kk",
    "speaker": "kk_M2"
  }' \
  --output output.wav

Stream raw signed 16-bit little-endian PCM:

curl -s http://localhost:8000/v1/audio/stream \
  -H 'Content-Type: application/json' \
  -d '{
    "text": "Real-time streaming speech is ready.",
    "language": "en",
    "speaker": "ljspeech_F1"
  }' | ffplay -f s16le -ar 22050 -ac 1 -

Use one server worker per loaded model instance. Scale with multiple service processes or containers when concurrent throughput is required.

Best Practices

• Pass language explicitly for every request.
• Use a speaker assigned to the selected language.
• Preserve punctuation to improve pauses and phrasing.
• Write numbers, dates, currencies, and abbreviations as words when exact pronunciation matters.
• Use proper Kazakh letters such as ә, ғ, қ, ң, ө, ұ, ү, һ, and і.
• Split long paragraphs into complete sentences and stream them in order.
• Reuse the loaded ONNX session; do not reload the model per request.
• Synthesize at 22,050 Hz and resample afterward if another rate is required.

Scope And Limitations

The model provides fixed speaker identities and does not clone arbitrary voices. Mixed-language text inside one request is not recommended; synthesize each language segment with its matching language and speaker. Pronunciation of unusual names, abbreviations, and ambiguous Russian word stress may require text normalization or explicit phonetic input.