SiliconFlow TTS

Generate narration audio from text using SiliconFlow Text-to-Speech API.

Use this skill when:

• You need voiceover or narration audio for a video
• You need standalone TTS assets before composing with Remotion/MoviePy
• You want to convert a script into reusable .mp3, .wav, .opus, or .pcm

Run

Do NOT set env vars inline (for example, SILICONFLOW_API_KEY=... python3 ...). The env var is already in the system environment; inline assignments break the exec permission check.

# Basic Chinese narration, saved under ./tmp/sf-tts-<ts>/speech.mp3
python3 ./skills/siliconflow-tts/scripts/tts.py --text "大家好，欢迎来到今天的视频。"

# Read text from a file
python3 ./skills/siliconflow-tts/scripts/tts.py \
  --text-file ./scripts/script.txt \
  --out-dir ./assets/audio

# Fragment workflow: read tts_requirement.md, extract voiceover/voice/speed,
# and output speech.mp3 + speech.json to ./fragments/01-hook/artifacts/
python3 ./skills/siliconflow-tts/scripts/tts.py ./fragments/01-hook/ --overwrite

# Select voice, format, and exact output path
python3 ./skills/siliconflow-tts/scripts/tts.py \
  --text "This is a demo voiceover." \
  --voice "fnlp/MOSS-TTSD-v0.5:benjamin" \
  --format wav \
  --sample-rate 44100 \
  --output ./assets/audio/demo.wav

Parameters

Recommended voices

Dialogue format

fnlp/MOSS-TTSD-v0.5 supports spoken dialogue scripts. Use speaker tags when writing multi-speaker dialogue:

[S1]Hello, how are you today?[S2]I'm doing great, thanks for asking!

Output

• Audio file: speech.<format> or the path set by --output
• Metadata file: speech.json beside the audio file, containing:
  • duration: audio duration in seconds (via ffprobe)
  • model, voice, format, text_chars, audio_bytes, file etc.

When used in the content-producer fragment workflow, pass the fragment directory directly. The script reads tts_requirement.md, extracts the ## 配音文案 / ## Voiceover Text section, reads voice/speed settings, and writes directly to the fragment's artifacts/ directory.

For tts_requirement.md, the script skips markdown headings, comments, and voice settings when synthesizing audio.

ASR Self-Check

After generating audio, the script automatically runs an ASR self-check (unless --no-asr-check is set):

1. Transcribes the generated audio via SiliconFlow ASR (TeleAI/TeleSpeechASR by default)
2. Compares transcription with the input text using Jaccard similarity
3. Threshold: 0.5 (50%) — based on testing, 50% Jaccard is sufficient for practical quality; higher thresholds caused excessive false negatives
4. Result printed as PASS or WARN; does not abort on failure

The ASR check calls /audio/transcriptions with multipart form fields file and model, matching SiliconFlow's transcription API. It requires SILICONFLOW_API_KEY; if not set, the check is silently skipped.

Environment Variables