dvcrn SenseAudio Polyphone TTS (多音字)
Precise pronunciation control for Chinese TTS via pinyin annotation. The dictionary parameter lets you override how specific characters are read — essential for polyphones (多音字) that the model might guess wrong.
The
dictionaryparameter only works with cloned voices and modelSenseAudio-TTS-1.5. System voices (male_0004_a etc.) do not support it.
Step 1: Scan for Polyphones
When the user provides text, scan it for these common polyphones and flag any that appear:
| Character | Readings | Context clues |
|---|---|---|
| 行 | háng (行业/银行/行列) / xíng (行走/行动/可行) | 银行、行长、行业 → háng |
| 干 | gān (干净/干燥) / gàn (干活/干部) | 干部、干活 → gàn |
| 量 | liáng (量体温/测量) / liàng (数量/重量) | 数量、质量 → liàng |
| 铺 | pū (铺床/铺路) / pù (店铺/铺子) | 店铺、铺面 → pù |
| 好 | hǎo (好的/很好) / hào (好奇/爱好) | 爱好、好学 → hào |
| 了 | le (吃了/来了) / liǎo (了解/了结) | 了解、了不起 → liǎo |
| 得 | de (跑得快) / dé (得到) / děi (得去) | 得到 → dé；必须 → děi |
| 地 | de (慢慢地) / dì (土地/地方) | 副词用法 → de |
| 的 | de (我的) / dí (的确) / dì (目的) | 目的、的确 → dì/dí |
| 着 | zhe (看着) / zháo (着火) / zhuó (着装) | 着火、着急 → zháo；着装 → zhuó |
| 长 | cháng (长度/很长) / zhǎng (成长/行长) | 行长、生长 → zhǎng |
| 重 | zhòng (重量/重要) / chóng (重复/重新) | 重复、重新 → chóng |
| 中 | zhōng (中间/中国) / zhòng (中奖/中毒) | 中奖、中毒 → zhòng |
| 还 | hái (还有/还是) / huán (还钱/归还) | 还钱、偿还 → huán |
| 发 | fā (发现/发展) / fà (头发/理发) | 头发、理发 → fà |
| 数 | shù (数字/数量) / shǔ (数数/数一数) | 数数、数落 → shǔ |
| 参 | cān (参加/参考) / shēn (人参/党参) | 人参、党参 → shēn |
| 差 | chā (差别/差距) / chà (差不多) / chāi (出差) | 出差 → chāi；差不多 → chà |
Show the user which polyphones were found and your best guess at the intended reading, then ask them to confirm or correct before synthesizing.
Example:
检测到多音字：
- "行" (第2个): 银行 → 建议读 háng [hang2] ✓ 还是 xíng [xing2]?
- "行" (第4个): 行长 → 建议读 zhǎng [zhang3] ✓ 还是 cháng [chang2]?
Step 2: Build the Dictionary
Convert confirmed readings into the dictionary array. Each entry covers one phrase containing the polyphone:
原文片段 → replacement 格式：在多音字前加 [pinyin]，其余字保持原样
Pinyin format: [声母韵母声调数字] — e.g., [hang2]、[xing2]、[zhang3]
Example:
- original:
银行行长 - replacement:
银[hang2]行[zhang3]长
Build the full dictionary array:
"dictionary": [
{"original": "银行行长", "replacement": "银[hang2]行[zhang3]长"},
{"original": "好奇心", "replacement": "[hao4]奇心"}
]
Each original should be a short phrase (3–8 chars) that uniquely identifies the occurrence in context. Avoid single-character originals — they may match unintended occurrences.
Step 3: Synthesize
The user must provide a cloned voice ID. If they don't have one, remind them that dictionary requires a cloned voice and suggest using the senseaudio-voice-cloner skill first.
curl -s -X POST https://api.senseaudio.cn/v1/t2a_v2 \
-H "Authorization: Bearer $SENSEAUDIO_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "SenseAudio-TTS-1.5",
"text": "<TEXT>",
"stream": false,
"voice_setting": {
"voice_id": "<CLONED_VOICE_ID>"
},
"audio_setting": {
"format": "mp3"
},
"dictionary": <DICTIONARY_ARRAY>
}' -o response.json
jq -r '.data.audio' response.json | xxd -r -p > output.mp3
Check base_resp.status_code == 0 before decoding.
Step 4: Iterate
After the user listens, they may find additional mispronunciations. Update the dictionary array and re-synthesize. Keep the previous response.json until the new one succeeds.
Report: file path, duration (jq '.extra_info.audio_length' response.json ms), character count, and which dictionary entries were applied.