Cheap TTS Python SDK

Generate multilingual speech with server-side text preparation and local ONNX inference. An active Cheap TTS subscription and a server secret key are required.

Requires Python 3.10–3.13. Create a project and server key in the Cheap TTS dashboard.

pip install cheap-tts

from cheap_tts import CheapTTS

with CheapTTS(secret_key="sk_live_your_project") as tts:
    result = tts.synthesize(
        text="Xin chào từ Cheap TTS.",
        voice="vi-entertainment",
    )
    result.save("audio.wav")

Async usage:

from cheap_tts import AsyncCheapTTS

async with AsyncCheapTTS(secret_key="sk_live_your_project") as tts:
    result = await tts.synthesize("Hello from Cheap TTS.", "en-narration")
    result.save("audio.wav")

API

• voices() returns the available voices.
• load_voice() downloads, validates, caches, and loads one voice.
• synthesize() prepares text through API v1 and performs ONNX inference locally.
• clear_cache() removes locally cached voice assets.
• close() releases the active model.

load_voice() and synthesize() accept on_progress and a cancellation object with an is_set() method, such as threading.Event.

from threading import Event

cancel = Event()
result = tts.synthesize(
    "A long document",
    "en-narration",
    rate=1.25,
    on_progress=lambda event: print(event.stage, event.progress),
    cancel_event=cancel,
)

Voice models are cached in ~/.cache/cheap-tts/models. Set cache_dir or CHEAP_TTS_CACHE_DIR to change the location.

The SDK sends operational health telemetry containing success/failure stage, latency, character count, voice ID, and audio duration. It never sends TTS text or credentials through telemetry. Pass telemetry=False to disable it.

API errors are raised as CheapTTSAPIError with stable status, code, and request_id fields.

• Full SDK documentation
• Package on PyPI
• Developer dashboard

Reading pauses

Pauses are left to the voice model by default. Pass pauses to insert real silence after a punctuation mark instead, in seconds (0-2):

result = tts.synthesize(
    "Xin chào từ Cheap TTS.",
    "vi-entertainment",
    pauses={"sentence": 0.4, "paragraph": 0.8},
)

Keys: sentence, comma, ellipsis, paragraph (a blank line). Every key is optional and defaults to 0, meaning the model's own pause — so omitting pauses keeps the audio you get today. Raising a value costs generation speed, because the audio has to be cut at that boundary for silence to follow it.