Voices and options
Options
The five settings that change how the model speaks, what it does when you leave them out, and the emotion tags it reads in the text.
The options#
Five, all optional. One you leave out isn't sent, and the model's own default applies.
normalizebooleanOptional; left out: on- Speak numbers, currency, dates, times, OTPs, phone numbers, URLs and emails as words in the request’s language.
falseturns it off: the text is spoken as written. temperaturenumberOptional; left out: the model's default- Sampling temperature, 0.1 to 1.5: lower is steadier, higher is more varied.
top_kintegerOptional; left out: the model's default- Sample from the k likeliest codes, 1 to 2048.
top_pnumberOptional; left out: the model's default- Nucleus sampling, above 0 and at most 1.
repetition_penaltynumberOptional; left out: the model's default- Penalty on recently repeated codes, 0 to 5.
A value outside its range is refused with bad_request, naming the option. GET /v1/models lists the options with their ranges.
Where they go#
The same five names in every client.
| Client | Where |
|---|---|
| LiveKit plugin | Keyword arguments of vaakini.TTS(...) and update_options(...) |
| Pipecat service | Fields of VaakiniTTSService.Settings(...) |
| Python SDK | Keyword arguments of speech.create(), speech.stream() and stream() |
| HTTP API | Top-level fields of the request body |
| WebSocket API | Inside options in start and warm |
from vaakini import Vaakiniclient = Vaakini() # reads VAAKINI_API_KEYaudio = client.speech.create( input="आपका बिल ₹1,250 है, जो 15 दिसंबर तक भरना है।", voice="Monika", language="hi", temperature=0.6, top_p=0.9,)with open("bill.mp3", "wb") as f: f.write(audio)curl https://api.vaakini.com/v1/audio/speech \ -H "Authorization: Bearer $VAAKINI_API_KEY" \ -H "Content-Type: application/json" \ -d '{ "model": "vaakini-tts-1", "voice": "Monika", "language": "hi", "input": "नमस्ते, आपका ऑर्डर कल पहुँच जाएगा। डिलीवरी से पहले आपको कॉल आएगा।", "response_format": "pcm", "temperature": 0.6 }' \ --output speech.pcmOn the WebSocket:
{ "type": "start", "segment_id": "reply-1", "model": "vaakini-tts-1", "voice": "Monika", "language": "hi", "sample_rate": 24000, "format": "pcm", "options": {"temperature": 0.6}}Numbers and symbols#
With normalize on, which it is unless you turn it off, what isn't a word is spoken as words.
Numbers, currency, dates, times, OTPs, phone numbers, URLs and emails are read in the request’s language: ₹1,250 in a Hindi request is spoken as an amount in rupees, in Hindi. Name the language when your text has them.
Turn it off (normalize: false) when your own code already writes everything out as words and you want the text spoken exactly as written.
Emotion tags#
Five tags, written in the text where the effect should happen.
| Tag | What it asks for |
|---|---|
| [laughs] | A laugh |
| [sighs] | A sigh |
| [whispers] | A whisper |
| [sad] | A sad tone |
| [excited] | An excited tone |
A tag is part of the text, not a setting, so it works the same in every client: write it inline, in square brackets, at the point in the sentence where it should take effect.
from vaakini import Vaakiniclient = Vaakini() # reads VAAKINI_API_KEYlines = { "hi": "[excited] अरे वाह, आपका ऑर्डर आज ही पहुँच जाएगा!", "en": "I am so sorry to hear that. [sighs] Let me see what I can do.",}for language, text in lines.items(): audio = client.speech.create(input=text, voice="Monika", language=language) with open(f"tags-{language}.mp3", "wb") as f: f.write(audio)- They are best effort. The model usually follows a tag, not always, and how strongly varies from take to take.
- They are counted. A tag’s characters are part of its sentence, brackets included:
[excited]adds 9 characters to the sentence’s count, and the space after it one more. - A tag belongs to the sentence it is written in. One written after a sentence’s end mark opens the next sentence. One at the very end of a text, after the last end mark, is sent as a sentence of its own: put a closing tag before the end mark instead.
What isn't an option#
- Speed. There is no
speed; a request that sends one is refused. - A seed. The model samples, so the same text can sound slightly different each time, and there is no setting that makes two takes identical. A lower
temperaturemakes them steadier. - Splitting. Text is always spoken sentence by sentence.
- The sample rate. Audio is 24000 Hz. Resample it on your side if you need another rate (the Pipecat pipeline’s output transport does it for you).