Introducing
High-accuracy speech recognition for Indian languages — phone calls at 8 kHz, meetings, voice notes and code-mixed speech, from a file or live.
Free in the Playground today · Public launch 1 October 2026
Speech recognition is easy when the audio is clean.Real-world speech is different.
That is exactly where we built Lark.
Building an AI caller, your own HeyPocket or Plaud, or your own NotebookLM? The part that listens is the part you shouldn't have to build. Lark hears Indian speech — phone lines, code-mix, several speakers — and hands your agent clean, structured text, so your time goes into the harness: the product, the prompts, the tools and the memory around the model.
You buildyour code
The harness
# one turn of a voice app
audio → Lark → text
text + prompt + tools → Osprey → reply
reply → Pica → voice
We runapi.minicrow.com
The senses
HearsLark · Lark LiveSpeech to text, from a file or a live call.This page
WatchesLark-VVideo to a summary, a timeline and every word.
ThinksOspreyHuman-like conversation and tool calling.
SpeaksPicaNatural voices in Hindi and English.
TalksOsprey LiveSees, hears, thinks and talks, in real time.One keyOne billPrepaid in rupeesCost in every response
Pocket, by HeyPocket, and Plaud Note and Plaud NotePin, by Plaud, are pocket-sized AI voice recorders: they capture calls, meetings and conversations and turn them into transcripts, summaries and action items.
The recorder is the easy part. The hard part is hearing — Indian languages, code-mix, phone calls, several people talking at once. That is Lark: send each recording and get back every word and who said it; then let Osprey write the summary and the action items.
Lark — the words and who said them
Osprey — a summary and the action itemsPocket is a product of HeyPocket; Plaud Note and Plaud NotePin are products of Plaud. They are named to describe the kind of device you can build — MiniCrow is not affiliated with them. The device in the videos is an illustration made for this page.
Voice agents that answer and make phone calls — receptionists, sales and support agents, appointment and payment reminders.
Live call · 01:12
Caller: “Mera order kab tak aayega?”
track_order(8841)
Agent: “Kal shaam 6 baje tak pahunch jaayega.”
Build it on MiniCrow
Lark Live hears the caller on an 8 kHz line. Osprey decides and calls your tools. Pica speaks the reply — or run all three on one socket with Osprey Live.
Lark LiveOsprey
PicaOsprey LiveLike NotebookLM, Google's AI research notebook — it answers from the sources you give it, such as documents, web pages, YouTube videos and audio, and turns them into podcast-style audio overviews.
Refunds are paid within 7 days of the return reaching us.2
Audio overview · 6:12
Build it on MiniCrow
Lark turns audio into text and Lark-V reads the videos. Embeddings & Reranking find the right passage, Osprey answers from it, and Pica reads it aloud.
Lark
Lark-VOspreyPicaNotebookLM is a product of Google. They are named only to describe the kind of app you can build — MiniCrow is not affiliated with them. The pictures are illustrations.
8 kHz
phone audio, zero accuracy lost
Measured twice against wideband: narrowing a recording to a phone line's 8 kHz cost nothing.
24
Indian languages
Hindi, Marathi, Tamil, Telugu, Bengali, Urdu and more — with English and nine international languages.
Live
turn by turn, on one WebSocket
A fast draft the moment a speaker stops, and a checked final right after. In preview.
S1 · S2
who said what, and when
Every turn with its speaker and timestamps — free when each speaker has their own channel.
अ · A
native script or romanised
Code-mix stays as spoken — “उद्या meeting आहे का?” — and the response says which script you got.
₹20
per hour of audio
From ₹20 an hour on lark-mini, the default — and lark-nano, the lightest, is coming soon. Billed by the second, prepaid, no subscription.
Declare the language and choose the alphabet: each language in its own script, or romanised the way people type in chat. English words stay in English, exactly as they were said.
24 Indian languages also live on Lark Live
And ten more, all live
Lark against the three speech-to-text APIs developers compare it with, on MiniCrow's current benchmark: 8 kHz telephone calls, 24–48 kHz recordings, and how much of the meaning survives. Best result in each row in bold.
Swipe the table to compare all four
| Benchmark | Larklark-mini, by MiniCrow Most accurate | SarvamSpeech to Text API | GoogleChirp 3, Cloud Speech-to-Text | ElevenLabsScribe v2 |
|---|---|---|---|---|
| Accuracy | ||||
| Word error rate, 8 kHz phone calls ↓1 | 20–25% | 18–27% | 22–30% | 22–28% |
| Word error rate, 24–48 kHz recordings ↓1 | 8–12% | 10–15% | 12–18% | 9–13% |
| Semantic meaning ↑2 | 95% | 91% | 90% | 90% |
| Price | ||||
| List price, per hour of audio3 | ₹20$0.208batch and real time (Lark Live) | ₹30$0.312 | ≈₹92$0.96 | ≈₹37$0.39 · real timebatch ≈₹21 · $0.22 |
MiniCrow's current benchmark measurements; results vary by language, accent, recording quality, dataset and evaluation methodology. Sarvam, Google and ElevenLabs are third-party providers named for comparison only; MiniCrow is not affiliated with them.
95% of the meaning, kept
More than just transcription: Lark understands Indian voices, accents and context.
Measured · 102 real calls
A third less, at the same accuracy
Lark mini sells at ₹20 per hour of audio. Sarvam's published rate is ₹30. On 102 real 8 kHz Marathi calls the two scored 82.3 and 80.8 — level, inside the noise. You are not trading accuracy for the price.
Press play: a support call, a voice note and a meeting, voiced by Pica and transcribed by lark-mini — word for word, as the audio reaches them.
Hear it in your language
Lark
Choose a scenario
8 kHz telephone audio · 21 languages
Lark's transcript
Spoken in Hindi by Pica's Aarav and Trisha; the words are lark-mini's own transcript, timed to the recording.
Hello, main aapke order ke baare mein call kar raha hoon.
Haan ji, order abhi tak deliver nahi hua, tracking bhi update nahi ho rahi.
Main check karta hoon, ek minute hold kariye.
Illustrative scenes.
LarkOspreyPicaLark handles the hearing. Osprey handles the reasoning. Pica handles the voice. And Osprey Live brings the real-time experience together — one platform instead of a dozen services stitched together.
A recording
One request: the language, the script, and who spoke when.
Python
import requests
r = requests.post(
"https://api.minicrow.com/v1/audio/transcriptions",
headers={"Authorization": "Bearer mc_YOUR_KEY"},
data={
"model": "lark-mini",
"language": "hi", # declared, never guessed
"script": "latin", # or "native"
"diarize": "true", # who spoke when
"speakers": "2",
},
files={"file": open("call.wav", "rb")},
)
print(r.json()["text"])A live call
Stream the line in; read every turn back as the caller speaks.
Python
# pip install "websockets>=14"
import asyncio, json, websockets
URL = ("wss://api.minicrow.com/v1/audio/transcriptions/live"
"?model=lark-mini&language=mr&script=latin"
"&encoding=mulaw&sample_rate=8000")
KEY = {"Authorization": "Bearer mc_YOUR_KEY"}
async def main():
async with websockets.connect(URL, additional_headers=KEY) as ws:
# stream the call's audio in as binary frames
async for message in ws:
event = json.loads(message)
if event["type"] == "transcript.final":
print(event["text"])
asyncio.run(main())Declare the language
Lark listens for the language you name — it never guesses.
Choose the alphabet
script=native or script=latin — English words stay in English.
Who spoke when
diarize=true adds speaker turns with start and end times, beside an unchanged transcript.
Two channels, two speakers
A stereo call with one party per channel is transcribed channel by channel — no labels to buy.
Names and terms
domain, vocabulary and abbreviations help Lark spell the names in your recordings.
The formats you have
WAV, OGG/Opus, MP3, M4A or WebM, up to 25 MB — billed by the second, read from the file's own header.
Cost in every response
In paise, on every call — and nothing charged when a transcription fails.
Live, over one WebSocket
μ-law, A-law or PCM at 8 or 16 kHz; a draft and a checked final for every turn.
Every field, limit and error is in the API reference.
Three tiers, one endpoint. Live prices per hour of audio, billed by the second, prepaid in rupees.
Coming soon
The lightest Lark.
The lightest Lark — for short clips and voice notes at volume.
₹10
per hour of audio · $0.104 an hour
The default
The one to build on.
The default — for calls, meetings and voice notes, in every language.
₹20
per hour of audio · $0.208 an hour
Lark Live Preview
Real-time speech to text over a WebSocket — ₹20 an hour ($0.208), billed per second of audio you send. Enabled per account.
Speaker labels
+₹3.50 an hour on any tier ($0.036) — not charged if the speaker pass fails, and free on a two-channel call.
| Model | For | Per hour |
|---|---|---|
| lark-nanocoming soon | Short clips, up to 5 minutes a request | ₹10$0.104 |
| lark-mini | The default — uploads and Lark Live | ₹20$0.208 |
| lark-large | No Indian language declared | ₹30$0.312 |
| lark-large | An Indian language declared | ₹45$0.468 |
| lark-large:max | The most accurate Lark | ₹60$0.624 |
| Speaker labels | diarize=true, on any tier; free on a two-channel call | +₹3.50$0.036 |
Live from our price book; dollars at ₹96.1. Every MiniCrow price
Production-grade speech recognitionshouldn't require production-grade pricing.
Public launch
1 October 2026.
Build › Deploy › Scale
High-quality speech intelligence, for every developer. Try Lark free in the Playground today — no account needed.
Lark is MiniCrow's speech-to-text (STT, or ASR) API for Indian languages and real-world audio — 8 kHz phone calls, Indian accents, code-mixed speech, several speakers and background noise. It turns speech into clean, structured text for AI voice agents, contact centres, meetings, transcription products and multilingual apps.
24 Indian languages on upload — Hindi, Marathi, Bengali, Tamil, Telugu, Gujarati, Kannada, Malayalam, Punjabi, Odia, Assamese, Urdu, Nepali, Sanskrit, Konkani, Maithili, Sindhi, Kashmiri, Dogri, Manipuri, Santali, Bodo, Bhojpuri, Rajasthani — and English, Spanish, French, German, Portuguese, Italian, Russian, Arabic, Japanese, Korean. Lark Live, the real-time stream, covers 11 of the Indian languages and all ten international ones. Declare the language with each request: Lark never guesses. Accuracy is measured most on Hindi and Marathi phone calls, and on read speech in 14 languages.
Yes — it is built for them. Telephone audio at 8 kHz was measured against wideband recordings twice, and narrowing the band cost zero accuracy both times. Send μ-law, A-law or PCM straight from the line; there is no need to upsample.
Yes. Send diarize=true — and speakers=2 when you know it — and every turn comes back with its speaker and its start and end times, beside an unchanged transcript. It adds ₹3.50 an hour to the tier's price, and nothing when the speaker pass fails or when a stereo call already carries one speaker per channel.
Yes, it is your choice: script=native writes each language in its own script, script=latin romanises it the way people type in chat. Code-mix stays as spoken — English words stay in English — and the response says which script you got.
Yes, with Lark Live: a WebSocket that answers every turn twice — a fast draft the moment the speaker stops, and a checked final right after. It costs ₹20 an hour, billed per second. Lark Live is in preview and enabled per account.
Speech translation is coming to Lark; it is not part of the API yet. Lark-V can already write a video's summary in another language — a Tamil video summarised in English, for example.
lark-mini ₹20 an hour, lark-large ₹30 (₹45 for an Indian language, ₹60 on lark-large:max) — $0.208 and $0.312–$0.624. lark-nano, the lightest tier, is coming soon. Billed per second of audio, prepaid in rupees, no subscription. Speaker labels add ₹3.50 an hour.
Lark is the hearing; you build the harness around it. For an AI recorder like HeyPocket's Pocket or Plaud Note, send each recording to Lark with speaker labels, then ask Osprey for the summary and the action items. For an AI caller, stream the call to Lark Live, let Osprey decide and call your tools, and answer with Pica — or use Osprey Live, which runs all three on one socket.
The request is OpenAI's shape — POST /v1/audio/transcriptions, multipart, with a file and a model — plus MiniCrow's own fields for language, script and speakers. Every response carries its cost.
MiniCrow launches Lark publicly on 1 October 2026. You can try it free in the Playground today.
Lark hears; the rest of MiniCrow thinks, speaks and watches. Press play: hear a real call, a voice — or watch Lark-V read a real clip.
Human-like conversational AI — fast, affordable LLMs for assistants, companions and agents, with tool calling.
Plan my Sunday — something relaxed.
Brunch at 11, a walk by the lake at 4. Shall I book the table?
₹6.9
per million input tokens · $0.072
Natural text to speech in Hindi and English, with expressive voices for assistants, IVR and narration.
“Welcome back, Riya! Your order is on its way.”
₹21
per hour of audio · $0.219
Video understanding — a summary, a seekable timeline and the spoken words, from a single clip.
Per clip
billed on what each clip used
Real-time multimodal AI that sees, hears, thinks and talks — with tool calling, for AI calling, companions, robots, glasses and cars.
Caller: “Do you deliver on Sundays?”
Osprey Live: “Yes — until 9 pm. Shall I book you a slot?”
Semantic search and RAG — embed your documents, rerank a shortlist, and give Osprey your own knowledge to answer from.
