silk mulberry 1.6

silk mulberry 1.6 is our multilingual voice. it speaks 22 indian languages, each written in its own script, in four named voices. you steer the delivery with a short style, accent and pace. it streams, so it’s a good fit for voice agents in any of those languages. it’s available over /v1/tts, /v1/tts/json, websocket streaming and the playground; the dashboard’s hosted agents and phone calls don’t offer it yet.

at a glance

field	value
model id	`mulberry-1.6`
language	22 indian languages, each in its own script. english words can stay in latin.
best for	voice agents, support lines and narration in indian languages
steering	an optional `description` of style, accent and pace, plus a named `speaker`
length	up to 2000 characters, and about 80 seconds of audio, per request

quickstart

one call, one .wav file:

curl -X POST https://silk-api.rumik.ai/v1/tts \
  -H "Authorization: Bearer rk_live_•••••••••" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "mulberry-1.6",
    "text": "வணக்கம், உங்கள் ஆர்டர் நாளை மாலைக்குள் வந்து சேரும்.",
    "description": "professional, Tamil accent, steady pace"
  }' \
  --output mulberry-1-6.wav

to receive opus, pcm, mulaw, alaw, or mp3, add audio_format to the request. see audio formats for exact values and examples. write your text in the language’s own script and put the delivery in description. there’s no language field: the text decides. the full request schema is in the api reference.

supported languages

write each language in its own script, and don’t romanise it. if your description names no accent, silk picks one from the script of the text.

language	write it in	default accent
assamese	bengali-assamese script	`Bengali accent`
bengali	bengali-assamese script	`Bengali accent`
bodo	devanagari	`Hindi accent`
dogri	devanagari	`Hindi accent`
english	latin	`Indian English accent`
gujarati	gujarati	`Indian English accent`
hindi	devanagari	`Hindi accent`
kannada	kannada	`Kannada accent`
kashmiri	perso-arabic	`Indian English accent`
konkani	devanagari	`Hindi accent`
maithili	devanagari	`Hindi accent`
malayalam	malayalam	`Indian English accent`
manipuri (meitei)	meetei mayek or bengali	`Bengali accent` in bengali script, else `Indian English accent`
marathi	devanagari	`Hindi accent`
nepali	devanagari	`Hindi accent`
odia	odia	`Indian English accent`
punjabi	gurmukhi	`Punjabi accent`
sanskrit	devanagari	`Hindi accent`
santali	devanagari, latin, bengali or odia	`Hindi accent` in devanagari, `Bengali accent` in bengali, else `Indian English accent`
tamil	tamil	`Tamil accent`
telugu	telugu	`Telugu accent`
urdu	perso-arabic	`Indian English accent`

name an accent in description to pick it yourself. the full list is in the prompting guide.

preset voices

pick one of four named voices with speaker. leave it out and you get ira.

voice	gender
`ira` `aisha` `siya` `zoya`	female

curl -X POST https://silk-api.rumik.ai/v1/tts \
  -H "Authorization: Bearer rk_live_•••••••••" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "mulberry-1.6",
    "text": "নমস্কার, আপনি কেমন আছেন? আজকের আবহাওয়াটা খুব সুন্দর।",
    "description": "happy, Bengali accent, steady pace",
    "speaker": "aisha"
  }' \
  --output aisha.wav

names are case-insensitive. send a name we don’t know and you get ira, not an error, so a typo sounds like the wrong voice rather than failing loudly. when you send speaker, /v1/tts/json echoes the voice it used, e.g. "Ira".

coming from mulberry 1.5? its numbered preset values, speaker_1 to speaker_4, map to ira, siya, aisha and zoya, in that order.

parameters

field	default	notes
`text`	n/a	required. up to 2000 characters, and about 80 seconds of speech. `<laugh>`, `<chuckle>` and `<sigh>` are performed where you place them; any other `<tag>` marker is removed before synthesis.
`description`	n/a	optional. `style, accent, pace`, e.g. `happy, Tamil accent, steady pace`. a part you leave out becomes `professional`, the script’s accent, or `fast pace`.
`speaker`	`ira`	optional. `ira`, `aisha`, `siya` or `zoya`.
`temperature`	`0.8`	sampling temperature.
`top_p`	`1.0`	nucleus sampling.
`top_k`	`30`	top-k sampling.
`max_new_tokens`	sized from the text	output length cap, set from your text: never below `2048`, up to `8192`. a larger value you send is kept, so you don’t need to set it.
`repetition_penalty`	n/a	accepted and ignored.

long text? one request speaks up to about 80 seconds of audio: roughly 1,200 english characters, or about 600 characters of indic-script text at a steady pace. split longer text into sentences. a generation that can’t finish comes back as an error instead of cut-off audio, and it isn’t charged: 502 GENERATION_TOO_LONG over http, or an error frame with code GENERATION_TOO_LONG over a websocket. any other failure, like the engine being unreachable, is GENERATION_ERROR.

how it differs from mulberry 1.5

silk mulberry 1.5 (mulberry) is unchanged. pick the model per request.

	silk mulberry 1.5	silk mulberry 1.6
model id	`mulberry`	`mulberry-1.6`
languages	hindi in devanagari, english in latin (code-mixed)	22 indian languages, each in its own script
voices	twelve named voices, or one built from `description`	four named voices, `ira` by default
`description`	required. one natural-language sentence	optional. `style, accent, pace`
delivery cues	`description` plus inline tags like `<laugh>`	`description` plus `<laugh>`, `<chuckle>` and `<sigh>`; other `<tag>` markers are dropped, `[tone]` markers aren’t supported
sampling defaults	`temperature` `0.6`, `top_p` `0.95`, `top_k` `50`	`temperature` `0.8`, `top_p` `1.0`, `top_k` `30`
`max_new_tokens`	`2048`. raise it for long lines	sized from your text
too long to finish	the audio comes back cut off	`GENERATION_TOO_LONG`, and not charged
endpoints	`/v1/tts`, `/v1/tts/json`, websocket streaming	the same, with the same frames
dashboard agents	yes, with phone calls	not yet

→ to write good descriptions, read the prompting guide.

documentation

sdk

integrations

cookbook

api reference

at a glance

quickstart

supported languages

preset voices

parameters

how it differs from mulberry 1.5

​at a glance

​quickstart

​supported languages

​preset voices

​parameters

​how it differs from mulberry 1.5

at a glance

quickstart

supported languages

preset voices

parameters

how it differs from mulberry 1.5