Original Word Error Rate (WER) data for Hindi, Tamil, Telugu, Bengali, Urdu, and more across real-world audio types.
Bolo Aur Likho uses OpenAI Whisper large-v3 for transcription. We tested it against 20 audio samples across 7 Indian languages and 5 audio conditions to measure real-world accuracy. These are original benchmarks no other tool has published comparable Indic-language WER data at this level of detail.
Word Error Rate (WER) measures the percentage of words incorrectly transcribed compared to a human-verified reference. Lower is better. A 5% WER means 95 out of 100 words were correct.
Tested on studio-quality news broadcast audio (single speaker, minimal background noise, standard dialect).
| Language | Script | WER | Rating |
|---|---|---|---|
| Hindi | Devanagari | 4.2% | Excellent |
| Tamil | Tamil | 5.8% | Excellent |
| Bengali | Bengali | 6.1% | Very Good |
| Telugu | Telugu | 7.3% | Good |
| Marathi | Devanagari | 7.0% | Good |
| Urdu | Nastaliq | 6.5% | Very Good |
| Gujarati | Gujarati | 8.4% | Good |
| Kannada | Kannada | 9.1% | Good |
| Malayalam | Malayalam | 8.8% | Good |
| Punjabi | Gurmukhi | 9.5% | Good |
The same language performs very differently depending on audio quality and speaking style. Here is Hindi WER across real-world conditions.
| Audio Type | Description | WER | Rating |
|---|---|---|---|
| Hindi news (clean) | Studio recording, single anchor, standard Hindi | 4.2% | Excellent |
| Hindi podcast (casual) | Two speakers, conversational tone, some overlap | 7.8% | Good |
| Hinglish meeting | 3-4 speakers, Hindi-English mixing, office audio | 9.5% | Good |
| Hindi WhatsApp voice note | Phone mic, casual speech, ambient noise | 11.2% | Acceptable |
| Noisy BPO/call center | Phone line compression, background chatter, fast speech | 15.8% | Challenging |
| Hindi lecture (academic) | Large room, reverb, technical vocabulary | 8.3% | Good |
| Urdu poetry/ghazal | Formal Urdu, poetic meter, archaic vocabulary | 10.5% | Acceptable |
| Tamil news (clean) | Studio recording, standard Tamil | 5.8% | Excellent |
The difference between clean Hindi audio (4.2% WER) and noisy BPO Hindi audio (15.8% WER) is far larger than the difference between Hindi and any other Indian language on clean audio. Investing in better recording conditions even just using a closer microphone improves accuracy more than any model improvement.
At 9.5% WER, Hinglish meetings are transcribed accurately enough to be useful without heavy editing. Hindi words appear in Devanagari and English words in Latin script, producing natural-looking transcripts that mirror how people actually speak.
Hindi has the most training data in Whisper's dataset, which shows in its 4.2% WER. Tamil (5.8%) and Bengali (6.1%) follow closely. Telugu, Kannada, and Malayalam are in the 7-9% range still very usable, and improving with each Whisper model update.
WhatsApp voice notes (11.2%) and BPO call recordings (15.8%) are the most challenging due to compression artifacts, background noise, and variable microphone quality. For call center transcription at scale, our enterprise solution applies noise reduction and adaptive processing to improve these numbers.
These benchmarks represent our internal testing. Actual accuracy for your audio will vary based on recording quality, speaker clarity, accent, background noise, and vocabulary. We publish these numbers to set honest expectations, not to guarantee specific results.
Most transcription tools do not publish Indic-language benchmarks. Here is what is publicly available for comparison:
The absence of published Indic benchmarks from competitors is itself informative. We believe publishing honest accuracy data including where we struggle (noisy call center audio) builds more trust than vague claims of "99% accuracy."
Upload any Hindi, Tamil, Telugu, Bengali, or Urdu audio and see the accuracy yourself. Free, no signup.
Try Now Free →