OpenSkillIndex← back to search
claude-code

Learn-skills.dev qwen-voice

Use Qwen (DashScope/百炼) for speech tasks: (1) ASR speech-to-text transcription of user audio/voice messages (Telegram .ogg opus, wav, mp3) using qwen3-asr-flash, optionally with coarse timestamps via chunking; (2) TTS text-to-speech voice reply using qwen3-tts-flash with selectable voice (default Cherry) and output as .ogg voice note for Telegram.

install
source · Clone the upstream repo
git clone https://github.com/NeverSight/learn-skills.dev
Claude Code · Install into ~/.claude/skills/
T=$(mktemp -d) && git clone --depth=1 https://github.com/NeverSight/learn-skills.dev "$T" && mkdir -p ~/.claude/skills && cp -r "$T/data/skills-md/ada20204/qwen-voice/qwen-voice" ~/.claude/skills/neversight-learn-skills-dev-qwen-voice && rm -rf "$T"
manifest: data/skills-md/ada20204/qwen-voice/qwen-voice/SKILL.md
source content

Qwen Voice (ASR + TTS)

Use the bundled scripts. Prefer environment variable 

DASHSCOPE_API_KEY
. If missing, scripts attempt to read it from 
~/.bashrc
.

ASR (speech → text)

Non-timestamp (default)

python3 skills/qwen-voice/scripts/qwen_asr.py --in /path/to/audio.ogg

With timestamps (chunk-based)

python3 skills/qwen-voice/scripts/qwen_asr.py --in /path/to/audio.ogg --timestamps --chunk-sec 3

Notes:

  • Timestamps are generated by fixed-length chunking (not word-level alignment).
  • Input audio is converted to mono 16kHz WAV before sending.

TTS (text → speech)

Preset voice (default: Cherry)

python3 skills/qwen-voice/scripts/qwen_tts.py --text '你好，我是 Pi。' --voice Cherry --out /tmp/out.ogg

Clone voice (create once, reuse)

  1. Create a voice profile from a sample audio:
python3 skills/qwen-voice/scripts/qwen_voice_clone.py --in ./voice_sample.ogg --name george --out work/qwen-voice/george.voice.json
  1. Use the cloned voice to synthesize:
python3 skills/qwen-voice/scripts/qwen_tts.py --text '你好，我是 George。' --voice-profile work/qwen-voice/george.voice.json --out /tmp/out.ogg

Notes:

  • .ogg
    output is Opus, suitable for Telegram voice messages.
  • Voice cloning uses DashScope customization endpoint + Qwen realtime TTS model.
  • Scripts use a local venv at 
    work/venv-dashscope
    (auto-created on first run).

Typical chat workflow

  • When user sends voice message/audio: run ASR and reply with the transcribed text.
  • When user explicitly asks for voice reply: run TTS and send the generated 
    .ogg
    as a voice note.