Asr

Interaction Flow

Step 0: Prerequisites Check

Before config setup, silently check the environment:

COLI_OK=$(which coli 2>/dev/null && echo yes || echo no)
FFMPEG_OK=$(which ffmpeg 2>/dev/null && echo yes || echo no)
MODELS_DIR="$HOME/.coli/models"
MODELS_OK=$([ -d "$MODELS_DIR" ] && ls "$MODELS_DIR" | grep -q sherpa && echo yes || echo no)

If coli is missing, stop here and do not proceed.

Step 0: Config Setup

Follow shared/config-pattern.md Step 0.

Initial defaults:

# 当前目录:
mkdir -p ".listenhub/asr"
echo '{"model":"sensevoice","polish":true}' > ".listenhub/asr/config.json"
CONFIG_PATH=".listenhub/asr/config.json"

# 全局:
mkdir -p "$HOME/.listenhub/asr"
echo '{"model":"sensevoice","polish":true}' > "$HOME/.listenhub/asr/config.json"
CONFIG_PATH="$HOME/.listenhub/asr/config.json"

Config summary display:

当前配置 (asr)：
  模型：sensevoice / whisper-tiny.en
  润色：开启 / 关闭

Setup Flow (first run or reconfigure)

Ask in order:

1. model: "默认使用哪个语音识别模型？"

   • "sensevoice（推荐）" — 支持中英日韩粤，可检测语言、情绪、音频事件
   • "whisper-tiny.en" — 仅英文

2. polish: "转录后由 AI 润色文本？（修正标点、去语气词、提升可读性）"

   • "是（推荐）" → polish: true
   • "否，保留原始转录" → polish: false

Save all answers at once after collecting them.

Step 1: Get Audio File

If the user hasn't provided a file path, ask:

"请提供要转录的音频文件路径。"

Verify the file exists before proceeding.

Step 2: Confirm

准备转录：

  文件：{filename}
  模型：{model}
  润色：{是 / 否}

继续？

Step 3: Transcribe

Run coli asr with JSON output (to get metadata):

coli asr -j --model {model} "{file}"

On first run, coli will automatically download the required model. This may take a moment — inform the user if models haven't been downloaded yet.

Parse the JSON result to extract text, lang, emotion, event, duration.

Step 4: Polish (if enabled)

If polish is true, take the raw text from the transcription result and rewrite it to fix punctuation, remove filler words, and improve readability. Preserve the original meaning and speaker intent. Do not summarize or paraphrase.

Step 5: Present Result

Display the transcript directly in the conversation:

转录完成

{transcript text}

─────────────────
语言：{lang} · 情绪：{emotion} · 时长：{duration}s

If polished, show the polished version with a note that it was AI-refined. Offer to show the raw original on request.

Step 6: Export as Markdown (optional)

After presenting the result, ask:

Question: "保存为 Markdown 文件到当前目录？"
Options:
  - "是" — save to current directory
  - "否" — done

If yes, write {audio-filename}-transcript.md to the current working directory (where the user is running Claude Code). The file should contain the transcript text (polished version if polish was enabled), with a front-matter header:

---