doubao-speech 0.1.0

Python client and CLI for Volcengine/ByteDance Doubao seed-tts-2.0 streaming TTS and bigmodel streaming ASR — full voice (speech-in + speech-out) toolkit.

doubao-speech

English | 中文

A production-minded Python client and CLI for Volcengine Doubao voice APIs — seed-tts-2.0 (text → speech) and bigmodel (speech → text) in a single package. Native-quality Chinese voices with emotion control, and streaming ASR with ITN and punctuation.

Why

doubao-speech is the first PyPI package that covers both directions of Volcengine's modern voice stack:

• Other Python TTS wrappers hit the older SAMI HTTP endpoint (no streaming, older voice quality).
• No published PyPI package speaks seed-tts-2.0 bidi-stream or the bigmodel ASR endpoint.

This package fills that gap with a clean, unified surface:

• synthesize() / transcribe() for text → speech / speech → text.
• A CLI that drops straight into agent frameworks (Hermes Agent, Dify, LangChain, n8n, …).
• Strict mypy on every public module.
• 95% unit test coverage, atomic output writes, proper credential redaction.
• Same credentials, same config file, same error hierarchy for both directions.

Install

pip install doubao-speech

# or with uv:
uv add doubao-speech

# CLI-only:
uv tool install doubao-speech

Quick start

Text → speech

from doubao_speech import synthesize

synthesize("你好，世界", "hello.mp3")

Speech → text

from doubao_speech import transcribe

text = transcribe("meeting.mp3")
print(text)

Async, both directions

from doubao_speech import synthesize_async, transcribe_async

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

transcript = await transcribe_async("interview.wav", enable_punc=True)

CLI

# TTS
doubao-speech say "你好" --out hello.mp3
doubao-speech say "好激动！" --voice zh-female-warm --speed 1.2 --out excited.mp3

# STT
doubao-speech transcribe meeting.mp3
doubao-speech transcribe voice-note.ogg --out transcript.txt
doubao-speech transcribe recording.wav --no-punctuation --sample-rate 16000

# Voice catalog
doubao-speech list-voices --lang zh

# Inspect effective config (tokens redacted)
doubao-speech config show

Credentials

Resolve order — first match wins:

1. Keyword arguments to synthesize(...) / transcribe(...)
2. Environment variables: VOLCENGINE_APP_ID, VOLCENGINE_ACCESS_TOKEN (also accepted as DOUBAO_APP_ID, DOUBAO_ACCESS_TOKEN)
3. ~/.doubao-speech/config.yaml
4. Built-in defaults

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

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

Credentials come from the Volcengine Speech console. You need seed-tts-2.0 activated for TTS and a bigmodel ASR resource enabled for STT (free tier suffices for testing).

Hermes Agent integration

Hermes Agent's declarative tts.providers.<name> command-type surface makes doubao-speech a one-liner:

# ~/.hermes/config.yaml
tts:
  provider: doubao
  providers:
    doubao:
      type: command
      command: 'doubao-speech say --text-file {input_path} --out {output_path}'
      output_format: mp3
      max_text_length: 1024
      timeout: 30

Any Hermes voice-out path now routes through Doubao seed-tts-2.0.

Hermes does not yet have a command-type STT provider; if you want Doubao ASR in Hermes today, use the bundled voice-volcengine plugin (separate install) for STT while doubao-speech handles TTS.

Audio format support

Direction	Input / Output	Notes
TTS	mp3 (default), wav, ogg, pcm	24 kHz by default
STT	wav, mp3, ogg, flac, raw PCM	Auto-detected from extension; requires ffmpeg for non-WAV inputs

For STT, non-WAV inputs are transcoded to PCM16 mono via ffmpeg at the target sample rate. Install ffmpeg once (brew install ffmpeg / apt install ffmpeg) and any format works.

Voices (TTS)

The CLI ships with curated aliases for common voices:

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; pass any raw speaker ID to voice= directly.

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
)

Supported emotions vary by speaker — check the Volcengine console.

STT features

• ITN (inverse text normalization): "一百二十三" → "123"
• Punctuation: Automatic commas, periods, question marks
• Disfluency removal (DDC): Strips "嗯", "啊", repeated syllables
• Utterance timestamps: Available on the async generator path
• Low-latency streaming: 200ms chunks by default; tune via --segment-ms

Disable any of the above with flags:

doubao-speech transcribe lecture.mp3 --no-itn --no-punctuation --no-ddc

Or in Python:

transcribe("lecture.mp3", enable_itn=False, enable_punc=False, enable_ddc=False)

Performance

• TTS: One synthesize call opens a fresh WebSocket and tears it down at the end. End-to-end latency for ~2 s of speech is ~750 ms on a healthy connection — network dominates.
• STT: Streaming starts returning partial transcripts within ~500 ms. Final transcript for a 10 s clip arrives in ~1.5-2 s total.
• import doubao_speech is ~3 ms — websockets and yaml are loaded lazily only when you actually call synthesize() / transcribe().

Error handling

All user-facing errors inherit from DoubaoSpeechError:

from doubao_speech import (
    DoubaoSpeechError, DoubaoConfigError,
    DoubaoAuthError, DoubaoAPIError, DoubaoTimeoutError,
    synthesize, transcribe,
)

try:
    transcribe("audio.mp3")
except DoubaoAuthError:
    ...  # rotate your token
except DoubaoTimeoutError:
    ...  # retry or check network
except DoubaoSpeechError as exc:
    ...  # catch-all

DoubaoTTSError is kept as a back-compat alias for users porting from the earlier doubao-tts package.

Security

• Access tokens are redacted in all logs and CLI output — see SECURITY.md for the exact policy.
• User text and transcribed content are not logged by default. Opt in with DOUBAO_SPEECH_TRACE_PAYLOADS=1 only for protocol debugging.
• ~/.doubao-speech/config.yaml is user-scoped; the shipped .gitignore excludes .env files.
• Report vulnerabilities: [e-mail] or a private GitHub security advisory.

Development

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

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 (TCP+TLS amortization), streaming callback API for partial transcripts, richer voice metadata sync.
• v0.3 — LangChain/LlamaIndex/Dify integration recipes.
• 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 and the bigmodel ASR endpoint.