Skip to navigation

Bulbul

View as Markdown

Bulbul v4 Flash is our latest text-to-speech model. It offers 200+ persona voices and uses the same API contract as Bulbul v3. Bulbul v3 remains the API default when model is omitted and offers 30+ classic speaker names. Both models accept up to 2500 characters per request.

At a Glance

Model IDbulbul:v4-flash (latest, recommended for new integrations), bulbul:v3 (API default when model is omitted)
What it doesConverts text into natural speech in 11 languages, with control over pace, pitch, loudness and temperature
Languages11 (10 Indian + English), full list
APIsREST, HTTP streaming, WebSocket
Input limits2,500 characters per REST request, 3,500 per HTTP stream request, 2,500 per WebSocket message; all limits
PricingPricing page
Best forVoice agents, IVR prompts, narration and audio content
Known limitationsSee below

Bulbul v4 Flash vs v3

Bulbul v4 FlashBulbul v3
StatusLatest, recommended for new integrationsAPI default when model omitted
Speaker voices200+ persona voices, organized by language and style (browse)37 named speakers
Default speakershubh_enhi_adsshubh
Languages11, with dedicated personas for 9 (details)11

To use Bulbul v4 Flash, set model="bulbul:v4-flash" and pick a speaker from the voice catalog. Everything else in your request stays the same.

Key Features

Large Voice Catalog

200+ persona voices on bulbul:v4-flash, organized by language and style (for example simran_en_customer, ishita_hi_devotional). bulbul:v3 offers 30+ named speakers such as Shubh, Ritu and Simran.

Indian Languages and Code-Mixing

Speak 11 languages (10 Indian languages and English) using BCP-47 language codes. Handles code-mixed text such as Hindi with English. The language code guides the text normalization applied before synthesis.

Voice Controls

Tune delivery with pace (0.5 to 2.0), pitch (-0.5 to 0.5), loudness (0.1 to 2.5) and temperature (0.01 to 1.0). All four work on both models.

Pronunciation Dictionaries

Pass a dict_id to control how brand names, acronyms and domain terms are spoken. Supported on both models.

Flexible Audio Output

Choose from wav, mp3, aac, flac, linear16, mulaw and alaw (opus at supported sample rates). Sample rates run from 8 kHz for telephony up to 48 kHz on REST, and up to 24 kHz on streaming.

REST, Streaming and WebSocket

Send up to 2,500 characters per REST request or WebSocket text message, and up to 3,500 characters per HTTP streaming request. The limits are the same on both models.

Language Support

Both models accept the same 11 languages, set with the language_code parameter. Bulbul v4 Flash has dedicated personas for nine of them. Malayalam and Odia work with any v4 persona, and the language code selects the language of the text.

LanguageCodebulbul:v3bulbul:v4-flash
Hindihi-INYesDedicated personas
English (Indian)en-INYesDedicated personas
Bengalibn-INYesDedicated personas
Gujaratigu-INYesDedicated personas
Kannadakn-INYesDedicated personas
Marathimr-INYesDedicated personas
Punjabipa-INYesDedicated personas
Tamilta-INYesDedicated personas
Telugute-INYesDedicated personas
Malayalamml-INYesAny persona
Odiaod-INYesAny persona

Bulbul reads code-mixed text such as Hindi with English words. Write Indic words in their native script (for example आपका order confirm हो गया है) for the best quality.

Available Speakers

Bulbul v3

Bulbul v3 has 37 named speakers. shubh is the default.

shubh, aditya, ritu, priya, neha, rahul, pooja, rohan, simran, kavya, amit, dev, ishita, shreya, ratan, varun, manan, sumit, roopa, kabir, aayan, ashutosh, advait, anand, tanya, tarun, sunny, mani, gokul, vijay, shruti, suhani, mohit, kavitha, rehan, soham, rupali

Bulbul v4 Flash

Bulbul v4 Flash has 200+ persona voices. A persona ID has three parts: a voice name, a language tag, and a style.

PartExamplesMeaning
Namesimran, ishita, shubhThe voice. One name can have several personas.
Language taghi, en, bn, ta, enhiThe language the persona is built for. enhi is English and Hindi code-mixed.
Styleconversational, customer, narration, ads, devotional, bankingThe delivery and use case.

For example, simran_en_customer is Simran speaking English in a customer-care style. Browse every persona, with audio samples, on the Voices page.

If you omit speaker on bulbul:v4-flash, the default is shubh_enhi_ads. Names from bulbul:v3 such as shubh are not valid on v4 Flash, and v4 persona IDs are not valid on v3. Using the wrong one returns a 400 that lists the speakers the model accepts.

Quickstart

Convert text to speech with Bulbul v4 Flash, recommended for new integrations.

from sarvamai import SarvamAI
from sarvamai.play import play, save
client = SarvamAI(
api_subscription_key="YOUR_SARVAM_API_KEY"
)
response = client.text_to_speech.convert(
text="Hello, how are you today?",
language_code="en-IN",
model="bulbul:v4-flash",
speaker="simran_en_customer",
)
# Play the audio
play(response)
# Save the response to a file
save(response, "output.wav")

The response is JSON with a request_id and an audios array. Each item is base64-encoded audio, WAV by default, so decode it before playing. The SDK’s save() helper does this for you.

Limits and Defaults

ItemValue
Max characters, REST (/text-to-speech)2,500
Max characters, HTTP stream (/text-to-speech/stream)3,500
Max characters, WebSocket2,500 per text message
pace0.5 to 2.0
pitch-0.5 to 0.5
loudness0.1 to 2.5
temperature0.01 to 1.0 (default 0.6)
speech_sample_rate, REST8000, 16000, 22050, 24000, 32000, 44100, 48000
speech_sample_rate, streaming8000, 16000, 22050, 24000
Default sample rate22050 Hz
output_audio_codec (REST)wav (default), mp3, aac, flac, linear16, mulaw, alaw, opus
Rate limitsSee Rate Limits

All values apply to both bulbul:v3 and bulbul:v4-flash. opus works at 8000, 16000, 24000 and 48000 Hz. At the default 22050 Hz it returns a 400, so set speech_sample_rate when you use it.

Known Limitations

LimitationDetailWorkaround
No SSML supportBulbul does not support SSML tags for fine-grained prosody controlUse pace for coarse control; split text at natural pause points
Romanised Indic input degrades qualityTransliterated input (e.g., "Aapka order confirm ho gaya hai") significantly reduces output qualityAlways use native script for Indic words (e.g., "आपका order confirm हो गया है")
High sample rates are REST-only32000, 44100 and 48000 are accepted on REST onlyUse 24000 Hz or lower for HTTP stream and WebSocket

Next Steps