Every captioning tool makes the same claim: “supports Indian languages.” A few are honest about what that means. Most pass English-trained models (Whisper, Deepgram, AssemblyAI) through a translation layer and ship the result. The transcription is technically there — but the script breaks, compounds split in the wrong places, and code-mixed speech comes back garbled.
We built the Aksharly Indic Engine on top of ASR trained first-party on Indian-language audio — not translated from English models. That changes four things in practice:
- Native script, no romanisation fallback. Telugu stays Telugu, Tamil stays Tamil. We ship the right font family for every script so what you see is what gets burned.
- Dravidian languages are first-class. Telugu, Tamil, Malayalam, Kannada are tuned, not translated. This is the biggest gap between us and English-first models.
- Code-mix is a mode, not a workaround. “Aaj meeting hai” comes back as that — Hindi + English in one line — not as two broken halves.
- Script-aware fonts. Our 6 font families render all 16 languages correctly. Competitor tools inflate font counts to the thousands but most of those fonts only support Latin.
Per-language — what we do well
Rather than publish a WER table with made-up decimals, here's what we can defend for each language, concretely:
Native-script output, no fallback romanisation
Noto Sans Telugu + Noto Serif Telugu bundled. Our Indic engine was trained on Telugu first-party audio, not translated from English models.
Correctly segments conjuncts and vowel marks
Tamil combines letters into visually dense clusters; generic ASR often splits them mid-word. Our engine keeps them whole, so captions align with how Tamil is written.
Handles long compound words without over-splitting
Malayalam is agglutinative — one “word” can be a full phrase. We keep compounds together instead of breaking every 2 characters like English-trained models.
Native-script output across all 16 varieties
Noto Sans Kannada bundled. Kannada and Telugu share aksharas but render differently — we keep the two straight.
Code-mix mode preserves English words inline
Everyday Hindi has English words threaded in (“aaj meeting hai”). Native-script mode writes those in Devanagari; code-mix mode keeps them in Latin letters where you actually said them.
Nukta and conjunct consonants rendered correctly
Noto Sans Bengali bundled. The font family handles the full conjunct set, not just the common 500 clusters Whisper covers.
Separated from Hindi, not folded in
Marathi shares Devanagari with Hindi but has distinct grammar and vocabulary. We label them separately so your captions sound like Marathi, not transliterated Hindi.
Right-to-left rendering, Nastaliq-style letterforms
Noto Nastaliq Urdu bundled. The caption burner respects RTL direction end-to-end, including word-level highlights.
What we won't pretend
- We don't publish a WER number yet. Our test set is still small and we would rather say “we don't know” than publish a decimal we can't stand behind. We're building a transparent benchmark suite — when it's ready, the dataset and numbers go on this page in full.
- Heavy regional accents can still trip us up. If your speaker shifts dialect mid-sentence (Telangana Telugu into Rayalaseema, for instance), we may miss a word. The editor lets you fix it in one tap.
- Background music over speech degrades accuracy. This is true of every ASR system. We don't pre-process to boost it artificially — what you upload is what we transcribe.
- Noise and clipping are not auto-cleaned. We don't de-noise your audio behind your back. If a word is unclear in the source, the caption will reflect that.
Why first-party Indian ASR matters
OpenAI's Whisper, which most “Indian language captioning” tools wrap, was trained on a corpus that is roughly 83% English. The remaining 17% is spread across 98 languages. Telugu, Tamil, Malayalam and Kannada together got well under 0.5% of training time. That shows up as:
- Dravidian vowel marks dropped or misplaced
- Compound words split at Latin-friendly boundaries
- Code-mixed speech routed to the dominant language, losing the other half
- Script fallbacks (Telugu rendered in Devanagari when the model is confused)
We built Aksharly specifically because, for Indian creators, those failure modes compound every single video. Our Indic engine is built to handle the exact patterns Whisper misses — and we iterate on it based on real Indian creator audio.
Test it yourself
Your first video is free with the full Studio Plus editor — upload a 60-second clip in any of our 16 languages and compare the output to anything else you're using. If we fall short on a specific language, email us the clip and we'll show you where in the pipeline it broke.