TeraTTSv2 ONNX

TeraTTSv2 is a self-contained ONNX Runtime text-to-speech release with selectable diffusion samplers, ten voice styles, Russian stress marking, and streamed audio output.

This release uses the clean English/Russian 25-second teacher and its matching eight-step CFG-3 distilled student.

Important — Russian stress is automatic. Text inside <ru>…</ru> receives stress markers automatically by default. Explicit + markers always win.

Important — cross-language prompts. When an English reference voice is speaking Russian, experiment with duration_scale below 1 (for example 0.8). It is usually a better starting point than the default 1.

Recommended voices: ru_f1 and ru_m5 are the preferred Russian voice prompts.

Installation

pip install -r requirements.txt

sounddevice is only required for direct speaker playback. On Linux, install the system PortAudio library if it is not already present.

Load with Transformers

from transformers import AutoModel

tts = AutoModel.from_pretrained(
    "TeraSpace/TeraTTSv2",
    trust_remote_code=True,
    provider="CPUExecutionProvider",
    threads=6
)

waveform = tts.generate_speech(
    "<ru>Привет от TeraTTS.</ru>",
    voice="ru_f1",
    duration_scale=1,
)
tts.save_wav("teratts.wav", waveform)

waveform is a mono float32 NumPy array at 44,100 Hz. save_wav writes standard signed-16-bit PCM WAV without an extra audio package.

To inspect the exact text passed to the encoder after number expansion, stress marking, and Unicode normalization, call tts.normalize_text(text).

Controls

Control	Values	Effect
voice	ru_f1 ★, ru_m5 ★, ru_f2, ru_m1, eng_f3, eng_f4_whisper, eng_f5, eng_m2_whisper, eng_m3, eng_m4	Selects a bundled precomputed voice style named after its reference audio. ★ marks the recommended Russian prompts.
duration_scale	Positive float, default 1	Higher values produce slower, longer speech.
diffusion_model	distilled (default), teacher	Distilled is faster; teacher supports adjustable CFG.
ruaccent_mode	full (default), dictionary	Full uses RUAccent neural ONNX graphs plus dictionaries; dictionary mode loads dictionaries only.

The default diffusion_model="distilled" is the fast eight-step sampler. To use the teacher sampler, choose it while loading:

teacher_tts = AutoModel.from_pretrained(
    "TeraSpace/TeraTTSv2",
    trust_remote_code=True,
    provider="CPUExecutionProvider",
    threads=6,
    diffusion_model="teacher",
)

guidance can be adjusted when generating with the teacher sampler. The distilled sampler has CFG 3 baked into its graph.

Language tags, numbers, and Russian stress

Language tags are required: wrap text in <en>…</en> or <ru>…</ru>. The runtime rejects untagged or unbalanced input with a tag-specific error. Before number expansion and stress marking, it inserts spaces after punctuation and between a number and a following word. Characters outside the model vocabulary are skipped with a runtime warning. Numbers inside language tags are expanded to words in the matching language before synthesis:

waveform = tts.generate_speech(
    "<ru>У меня 21 яблоко.</ru> <en>I have 42 apples.</en>",
    voice="ru_f1",
    duration_scale=1,
)

Russian text is automatically stress-marked by the bundled RUAccent-derived runtime. Manual + markers remain authoritative. For a lower-memory, deterministic dictionary-only path, choose the mode while loading:

dictionary_tts = AutoModel.from_pretrained(
    "TeraSpace/TeraTTSv2",
    trust_remote_code=True,
    provider="CPUExecutionProvider",
    threads=6,
    ruaccent_mode="dictionary",
)

Dictionary mode does not load RUAccent neural ONNX graphs. It marks known words and applies deterministic ё replacements, while unknown words and ambiguous homographs are left unchanged. Set russian_stress=False to disable automatic Russian stress processing entirely.

When using an English voice such as eng_f3 for Russian text, start by trying duration_scale=0.8 and adjust by ear:

waveform = tts.generate_speech(
    "<ru>Это русский текст английским голосом.</ru>",
    voice="eng_f3",
    duration_scale=0.8,
)

Stream audio

for chunk in tts.generate_speech_stream(
    "<en>Streaming speech is ready.</en>",
    voice="eng_f3",
    duration_scale=1,
):
    # Send float32 mono chunks (44,100 Hz) to a player or network client.
    consume(chunk)

The remote code loads only the selected sampler graph plus shared ONNX graphs. For security, pin a specific Hub commit when using trust_remote_code=True.

Attribution

The local Russian stress annotator and its assets are adapted from RUAccent, Copyright 2026 Denis Petrov, under the MIT License. See RUACCENT_NOTICE.txt.