跳至正文
OpenAgentSkill
浏览已开放 · 目录记录日期 2026-11-21 · 自动更新和账户服务暂时关闭。
开发与测试

svara-multilingual

Speak Indian, African, Asian and other non-English text naturally with Svara TTS Turbo, which covers 82 languages and switches automatically inside mixed-script text such as Hinglish (Hindi plus English). Use for Hindi, Bengali, Tamil, Telugu, Marathi, Gujarati, Kannada, Malayalam, Punjabi, Odia, Urdu, Bhojpuri, Arabic, Swahili, Yoruba, Japanese and other language TTS; for code-mixed text; for forcing a language or normalising numbers and dates; and for fixing the pronunciation of brand names and acronyms with pronunciation dictionaries.

Kenpath Labs · 已有 AI 审核记录

安装与使用

运行命令前先阅读来源与依赖说明。复制命令不会自动安装或执行。

npx skills add kenpath-labs/svara-python --skill svara-multilingual

技能说明


name: svara-multilingual description: Speak Indian, African, Asian and other non-English text naturally with Svara TTS Turbo, which covers 82 languages and switches automatically inside mixed-script text such as Hinglish (Hindi plus English). Use for Hindi, Bengali, Tamil, Telugu, Marathi, Gujarati, Kannada, Malayalam, Punjabi, Odia, Urdu, Bhojpuri, Arabic, Swahili, Yoruba, Japanese and other language TTS; for code-mixed text; for forcing a language or normalising numbers and dates; and for fixing the pronunciation of brand names and acronyms with pronunciation dictionaries. license: Apache-2.0 compatibility: Needs SVARA_API_KEY and network access to api.kenpathlabs.com. Python 3.9+ with pip install svara-voice, or any HTTP client. metadata: author: Kenpath Labs version: "1.0" homepage: https://docs.kenpathlabs.com/pronunciation

Multilingual speech with Svara TTS Turbo

Svara speaks 82 languages:

  • 28 from the Indian subcontinent, including Bhojpuri, Maithili, Santali, Konkani, Dogri and Manipuri
  • 30 African
  • 10 European
  • 9 Asian, including Chinese, Japanese and Korean
  • 5 from the Middle East

The full table of codes is in references/languages.md.

Write text the way people write it

Svara reads the script of the input and switches language on its own, mid-sentence if needed. Do not split mixed text into per-language requests. Do not transliterate it, and do not add markup (there is no SSML):

from svara import Svara
client = Svara()

client.speech.save("order.mp3", voice="sv_fhn6tfve",
    input="आपका order ship हो गया है, और कल तक deliver हो जाएगा।")      # Hinglish, one call
client.speech.save("ta.mp3", voice="sv_rum9hcg9",
    input="வணக்கம்! Your appointment is confirmed.")                  # Tamil + English

Write each language in its own script: detection works from the script, so Hindi typed in Latin letters ("aapka order") gives the model no signal that it is Hindi. If the source is romanised, have the LLM that writes the reply output Devanagari (or the language's own script) instead.

When to pass language=

Language detection is automatic, so language is optional. Pass it when you want:

  1. Number, date, currency and unit normalisation in that language. For example, "₹2,500 on 15/08" is read as Hindi words when language="hi" is set. Normalisation is on only when a language is given.
  2. To settle an ambiguous script. Urdu and Arabic share a script, so do Hindi, Marathi and Nepali, and Chinese and Japanese share characters.
  3. A low-resource language whose name you know.

The argument accepts an ISO-639-1 code (hi), an ISO-639-3 code (hin), a name (hindi), an alias (mandarin, naija) or a BCP-47 tag (hi-IN, zh-CN). On the wire the field is lang:

client.speech.create(input="कुल राशि ₹2,500 है।", voice="sv_84sb2v3w", language="hi")
curl -s https://api.kenpathlabs.com/v1/languages        # live list, no key needed

normalize=False turns normalisation off when the text is already written the way it should be spoken.

Choosing a voice for a language

Every voice speaks every language. A voice native to the language gives the most natural accent. Find one with the svara-voices skill:

python ../svara-voices/scripts/find_voices.py --language bn -q A

There are native voices for English, Hindi, Arabic, Portuguese, German, Gujarati, Tamil, Indonesian, Bengali, Telugu, Punjabi, Turkish, Spanish, Russian, French, Italian, Korean, Urdu, Kannada, Malayalam, Marathi, Vietnamese and about 40 more. Japanese and Chinese have no native voice: pick a voice by style and pass language="ja" or "zh".

Pronunciation dictionaries: brand names, acronyms, jargon

Rules are respellings, not IPA. Write each one the way the word should sound, in whatever script gets that sound across:

from svara import PronunciationRule

d = client.pronunciation_dictionaries.create_from_rules(name="acme-brand", rules=[
    PronunciationRule("SQL", "sequel"),
    PronunciationRule("Kenpath", "Ken path"),
    PronunciationRule("HDFC", "एच डी एफ सी", only_languages=["hi"]),   # only on Hindi requests
])
client.speech.create(input="HDFC का SQL dashboard", voice="sv_84sb2v3w",
                     language="hi", pronunciation_dictionary_id=d.id)
  • pronunciation_dictionary_id works on create, stream, stream_input, prepare, the timestamp calls, and the LiveKit and Pipecat integrations.
  • Rule options are case_sensitive, word_boundaries (default on), only_languages and except_languages.
  • An unknown id is not an error: the global rules apply instead, and the response header x-svara-dictionary: miss reports it (the SDK warns). Copy ids from the console at https://platform.kenpathlabs.com.
  • Creation is all-or-nothing. A duplicate name returns 409. A full plan returns 403.
  • You can list dictionaries in code, but deletion is done in the console.

REST: POST /v1/pronunciation-dictionaries/add-from-rules with {"name":"acme-brand","rules":[{"text":"SQL","pronunciation":"sequel"}]}.

Tips

  • The Devanagari danda । and the CJK 。 end a sentence the same way . does.
  • Use speed (0.7–1.5) to slow dense content such as instructions. Do not insert extra punctuation to slow it down.
  • For a voice agent that answers in the user's language, keep a single voice and let the LLM reply in that language. Svara follows the script, so the voice stays the same.