doubao-tts 0.1.0

Python client and CLI for Volcengine/ByteDance Doubao seed-tts-2.0 bidirectional streaming TTS.

doubao-tts

English | 中文

A small, production-minded Python client and CLI for Volcengine Doubao seed-tts-2.0 bidirectional streaming TTS — native-quality Chinese voices with emotion control, ready for agents, scripts, and serving pipelines.

Why

doubao-tts is the first PyPI package targeting Volcengine's seed-tts-2.0 bidirectional-streaming endpoint. Existing Python TTS wrappers either:

• hit the older SAMI HTTP endpoint (no streaming, older voice quality), or
• aren't published to PyPI at all.

This package fills that gap with:

• A clean synthesize(text, out_path) interface
• A CLI that drops straight into agent frameworks (Hermes, Dify, LangChain, n8n, …)
• Strict mypy on every public module
• 95% unit test coverage, atomic output writes, proper credential redaction

Install

pip install doubao-tts

# or with uv:
uv add doubao-tts

# CLI-only, installed as a standalone tool:
uv tool install doubao-tts

Quick start

Python

from doubao_tts import synthesize

synthesize("你好，世界", "hello.mp3")

Async

from doubao_tts import synthesize_async

await synthesize_async(
    "Hello from Doubao seed-tts-2.0!",
    "hello.mp3",
    voice="en-female-assistant",
    speed=1.1,
)

CLI

# simplest
doubao-tts say "你好" --out hello.mp3

# pick a voice + adjust speed
doubao-tts say "好激动！" --voice zh-female-warm --speed 1.2 --out excited.mp3

# read from a file
doubao-tts say --text-file script.txt --out narration.mp3

# browse available voices
doubao-tts list-voices --lang zh

# inspect resolved config (tokens are redacted)
doubao-tts config show

Credentials

Credentials resolve in this order — first match wins:

1. Keyword arguments to synthesize(...)
2. Environment variables: VOLCENGINE_APP_ID, VOLCENGINE_ACCESS_TOKEN (also accepted as DOUBAO_APP_ID, DOUBAO_ACCESS_TOKEN)
3. ~/.doubao-tts/config.yaml
4. Built-in defaults (speaker, audio format, sample rate)

Example ~/.doubao-tts/config.yaml:

app_id: "1234567890"
access_token: "volc_...."
speaker: zh_female_vv_uranus_bigtts
audio_format: mp3
sample_rate: 24000

Get your app ID and access token from the Volcengine Speech console. You need the seed-tts-2.0 product activated on your account.

Integration: Hermes Agent

Hermes Agent v0.x+ supports declarative TTS command providers via its tts.providers.<name> config block. Plug doubao-tts in:

# ~/.hermes/config.yaml
tts:
  provider: doubao
  providers:
    doubao:
      type: command
      command: 'doubao-tts say --text-file {input_path} --out {output_path}'

That's it. Any Hermes voice-out path now routes through Doubao seed-tts-2.0.

Voices

The CLI ships with a curated alias catalogue:

Alias	Language	Gender	Style
zh-female-warm (default)	zh-CN	female	warm, conversational
zh-female-reporter	zh-CN	female	crisp, news-reporter
zh-male-warm	zh-CN	male	warm, narrator
zh-male-energetic	zh-CN	male	energetic host
en-female-assistant	en-US	female	assistant, neutral
en-male-assistant	en-US	male	assistant, neutral

Volcengine publishes hundreds more speaker IDs. You can pass any raw speaker ID to voice= directly — aliases are a convenience, not a gate.

Emotion control

seed-tts-2.0 supports per-utterance emotion tags:

synthesize(
    "好激动，我终于做到了！",
    "out.mp3",
    emotion="excited",
    emotion_scale=4.0,  # 0-5; higher = more intense
)

Model-supported emotions vary by voice; consult the Volcengine console for the up-to-date list per speaker.

Performance notes

• One synthesize() call opens a fresh WebSocket and tears it down at the end. End-to-end latency to a 24 kHz MP3 of ~2 seconds of speech is ~750 ms on a healthy connection — network dominates.
• The seed-tts-2.0 bidi-stream session currently accepts one synthesis per session (empirically verified); connection reuse saves only TCP+TLS setup (~180 ms / call). A daemon mode with connection pooling is planned for a future release, but most users don't need it.
• import doubao_tts is cheap — ~3 ms — because websockets and yaml are only imported on first synthesis call.

Error handling

All user-facing errors inherit from DoubaoTTSError:

from doubao_tts import (
    DoubaoTTSError, DoubaoConfigError,
    DoubaoAuthError, DoubaoAPIError, DoubaoTimeoutError,
    synthesize,
)

try:
    synthesize("你好", "out.mp3")
except DoubaoAuthError:
    ...  # rotate your token
except DoubaoTimeoutError:
    ...  # retry or check network
except DoubaoTTSError as exc:
    ...  # catch-all

Security

• Access tokens are redacted in all logs and CLI output — see SECURITY.md for the exact policy.
• User text is not logged by default. To troubleshoot protocol issues, opt in with DOUBAO_TTS_TRACE_PAYLOADS=1.
• ~/.doubao-tts/config.yaml is user-scoped; the shipped .gitignore excludes .env files at the project level.
• Vulnerability reports: [e-mail] or a private GitHub security advisory.

Development

git clone https://github.com/Hypnus-Yuan/doubao-tts.git
cd doubao-tts

uv sync --all-extras --group dev
uv run pre-commit install
uv run pytest

See CONTRIBUTING.md for the full workflow.

Roadmap

• v0.2 — connection-reuse daemon (saves ~180 ms / call on chained requests), streaming callback API, richer voice metadata.
• v0.3 — integration recipes for LangChain, LlamaIndex, Dify.
• v1.0 — API frozen, semver guarantees.

License

MIT — see LICENSE.

Credits

Protocol framing extracted and hardened from Hermes Agent community work. Thanks to the Volcengine Speech team for the seed-tts-2.0 bidirectional-streaming API.