A beautiful video can lose you in one mispronounced word. Hasyam is the on-device pronunciation layer inside crushedwithit—built for the names, languages and mixed sentences that make a voice your own.
You might write Hindi in Latin letters, keep an English phrase in the middle, then end on a name from home. Each part needs a different decision.
Intelligence at the handoff.
Hasyam combines bundled vocabulary, context rules and your personal pronunciation choices. It prepares the spoken form on your device, so the video model receives a more explicit version of what you meant to say.
Your original writing stays available for captions. You can inspect “Said as”, correct a name, or keep a word in English before spending credits on a render.
A compact vocabulary-and-rules engine prepares pronunciation locally. Your selected online video model creates the performance, with the cost and consent shown in the app.
ONE LINE. TWO JOBS.
YOU WRITE · CAPTIONS KEEP THIS
Bhai, main client ko bola ki lunch pe milte hain.
↓
PREPARED DIALOGUE · FOR THE VIDEO MODEL
भाई, मैं client को बोला कि lunch पे मिलते हैं।
Illustrative text example. Highlighted English words stay in Latin letters. Final generated speech still needs listening and review.
01 / ON DEVICE
Read the context.
Language, vocabulary and surrounding words guide the choice.
02 / YOUR CONTROL
Prepare. Inspect. Correct.
Native script where useful. Personal spellings where they matter. Uncertain words flagged.
03 / ONLINE RENDER
Give the model direction.
Prepared dialogue and delivery guidance go to the selected video service when you create.
04 / BACK TO YOU
Listen. Keep. Refine.
Review the result. Save your pronunciation choice for the next line or fix a supported part.
THE DATA BEHIND THE DETAILS
Local vocabulary. Visible evidence.
Six native-language dictionaries, each with 50,000 Roman spellings. Multiple native forms help represent the ambiguity that a single spelling can hide.
BUNDLED VOCABULARY
Roman → native pairs
Unique spelling-to-script pairs in each dictionary.
075,000150,000 pairs
Hindi / Hinglish130,284
Telugu114,073
Tamil106,473
Kannada115,975
Marathi110,129
Punjabi101,909
Hindi and Hinglish share one dictionary. English has a separate preservation list and is excluded from these totals. Larger lists do not establish better speech quality.
INTERNAL TEST FIXTURES
Reference lines
Written lines paired with expected prepared text.
0200400 lines
Hinglish340
Telugu123
Tamil118
Kannada118
Marathi119
Punjabi161
Counts show test-set size, not passed tests or audio accuracy. The Telugu, Tamil, Kannada, Marathi and Punjabi sets still need native-speaker review.
How these figures were measured
We counted distinct Roman/native pairs in the bundled resources, including compressed files, and two-column reference lines in the internal fixtures. Comments and blank or invalid rows are excluded. These counts were taken on 29 September 2026.
They describe vocabulary size and evaluation coverage. They are not training-parameter counts or evidence of superiority over another model. We have not completed a controlled public audio comparison with other providers.
The sentence has to sound right. The chart is only the beginning.
Our next evaluation step is a matched audio study: the same scripts, voices and settings, with and without prepared pronunciation, assessed by native speakers. We want to measure word pronunciation, names, English preservation and listening preference—and publish the sample sizes and limitations with the results.
ROOTED IN INDIA’S OPEN LANGUAGE RESEARCH
Built with open knowledge. Focused on your next line.
AI4Bharat’s work at IIT Madras is making Indian-language technology possible at a different scale. Hasyam builds on one practical part of that foundation: connecting the way a word is typed to the script that helps express how it is said.
IN HASYAM TODAY · DATA SOURCE
AI4Bharat Aksharantar.
The published corpus contains 26 million transliteration pairs. Our app bundles a selected six-language vocabulary derived from it, merged with curated entries and combined with context rules and personal corrections.
Our contribution is the preparation workflow: preserve English where intended, handle mixed lines, surface uncertain choices and carry your correction into the next video request.
26Mpairs in AI4Bharat’s published corpus↓678,843pairs in our bundled dictionaries, including curated entries
Where the research can take us.
The following are AI4Bharat resources we can study for future work. They are not integrated into Hasyam today, and their scale is not a claim about our training.
FUTURE RESEARCH · SPEECH
Listen to the way India speaks.
IndicVoices reports 23,700 hours across 22 languages and more than 400 districts. IndicConformer-600M is AI4Bharat’s 600-million-parameter speech recogniser. This work offers a foundation to investigate broader speech evaluation.
FUTURE RESEARCH · MEANING
Carry meaning across languages.
BPCC and IndicTrans2 connect roughly 230 million text pairs with translation research across 22 scheduled languages. Samanantar contributes a separate English–Indic corpus of 49.7 million pairs. These resources open questions about future translation workflows.
FUTURE RESEARCH · EXPRESSION
Go beyond saying the words.
Rasa brings expressive speech, conversational styles and emotion into multilingual text-to-speech research. Its published release covers 44 speaker-language pairs across 22 languages—a useful direction for studying delivery.
More from the AI4Bharat research library
IndicXlit explores neural transliteration. MILU evaluates multilingual language understanding. IndicBERT-v3 provides multilingual encoder models, while IndicDLP addresses document layout. These are separate research resources, not components of the current app.
Source descriptions checked on 29 September 2026. AI4Bharat’s datasets and models are credited to their creators; mentioning them does not imply a partnership or endorsement.
FROM A WORD TO A WHOLE PERFORMANCE
Longer videos. Finer control.
Pronunciation is one part of the performance. crushedwithit carries it into a workflow for making, reviewing and improving longer Personal Double videos.
PLAN THE LINE
Script and cuts.
Review prepared dialogue, scene details and the price before rendering.
BUILD THE PERFORMANCE
One part at a time.
Generate short parts with continuity guidance, then join the finished parts into one video.
KEEP WHAT WORKS
Fix the part.
Compare takes, remake a selected part or correct pronunciation where editing is supported.
The pronunciation layer runs on device. Personal corrections can sync through your private iCloud. Online rendering sends the approved material to the service disclosed in the app.
Put a language preparation step in front of your video model. The Hasyam API preview accepts a script, a language and your pronunciation choices, then returns prepared dialogue, original captions and words to review. Your application stays in control of rendering.
Agency production.
Send your brand’s pronunciation guide with each request, for names, places and recurring campaign language.
Video products.
A preparation step between a creator’s script and the model that performs it.
Evaluation partners.
Native-speaker review and matched samples that show where the approach helps—and where it still needs work.
A working, key-protected preview on Cloudflare. Eight language choices. 8,000 characters per request. Access is arranged by email; self-service signup, billing and a business SLA are not available. This server preview is separate from the app’s on-device engine and does not generate video.
HASYAM · BY CRUSHEDWITHIT
Keep what makes you sound like you.
Eight language choices. Personal pronunciation. Your words at the centre.