Retroflex consonants expose the recognition gap first. An agent that handles only one English variety cannot ship in India. Before Ringg reached our early-access teams, we had to address what speech systems trained mainly on North American data often miss: Indian English is a group of accents shaped by phonological patterns from more than a dozen regional languages. Each changes English vowels and consonants differently.
Here is what we changed in recognition, before running a single production call.
The Phonetic Challenge
ASR models trained on Librispeech or similar English data work well with General American or standard British English. Word error rates there usually land at 5 to 8 percent. With Indian English, the same models can reach 18 to 25 percent, depending on regional background. This is patterned degradation, not random noise.
Retroflex consonants cause the largest error share. In Tamil, Telugu, Kannada, and Hindi, the tongue tip curls back to make /t/ and /d/. Speakers carry that habit into English, so "Doctor" may use retroflex /t/ rather than alveolar /t/. Models trained on American speech can assign another phoneme, sending errors through the transcript.
Vowel reduction creates another major problem. Indian English often keeps unstressed vowels instead of turning them into schwa as American English does. "Computer" retains four vowels, and "Customer" does not become the American /k-uh-st-mur/ pattern. A recogniser expecting schwa hears unfamiliar sounds and substitutes unstressed syllables throughout the call.
Prosody differs too. Across several Indian language families, yes-no questions often end with rising intonation. If recognition treats that rise as an unfinished utterance, it gets turn endings wrong and disrupts call timing. The agent may speak before the caller finishes, one of the most jarring caller experiences.
What Errors Break During Live Calls
This failure is tangible. Mishearing a name in the opening puts the call off course at once. Early transcripts showed agents calling a caller "Rajiv" when the name was "Ragesh", creating an awkward pause that callers notice even when silent. In a 90-second qualification call, that trust loss is immediate and difficult to repair.
Intent errors matter more. Transcribing "I want to know more" as "I don't know" reverses the signal. "My budget is flexible" should lead to a different question than "my budget is fixed." When unstressed middle syllables degrade, the intent engine receives faulty input each turn.
Outbound calls magnify the issue because the caller did not request them and has little patience for confusion. An inbound support caller may patiently repeat themselves. On an outbound qualification call, the same error can earn a hang-up.
Changes to the Pipeline
We adapted three areas: the acoustic model, language-model rescoring, and end-of-speech detection.
For acoustics, we added curated Indian English recordings from four regional backgrounds: Hindi-belt speakers, South India (Tamil and Kannada speakers), Western India (Marathi and Gujarati speakers), and Bengali speakers. We avoided one "Indian English" model, which would be mediocre across all four variants. Instead, lightweight acoustic adapters load at call time using a confidence signal from the first two turns.
The rescoring layer resolves named entities. Indian names often resemble common English words, while a general language model assigns them low probability. We load a per-call name lexicon from CRM data, giving the caller's first and last name strong priors during n-best rescoring. This alone cut caller-name errors by roughly half in our internal test set.
End-of-speech detection needed separate treatment. Many Indian English speakers keep pitch level or rising at sentence end, unlike American speakers. Our old turn logic cut in early because it expected a falling final cue. We adjusted the turn-detection energy threshold and added a 150ms trailing-silence window tuned to Indian English. False barge-ins fell without noticeable added turn latency.
Honest Limits
This is not solved. Retroflex and alveolar confusion still causes substitutions for some names beginning with /t/ or /d/ when spoken by Tamil speakers. Our word error rate for that group is around 12 percent, below the untreated baseline but still too high. The adapter warm-up, the first two turns, also has higher error rates, which matters on short calls.
Our data also leans toward urban, educated Indian English, whose phonological range is narrower than the country's full range. A first-generation English speaker from a smaller town will face more errors than a metro-area speaker. That reflects a real current limitation.
This work targets the ASR layer only. It does not address how Indian English phonology affects conversation structure, including politeness markers, question tags, or more indirect phrasing in some regional styles. Those are script and conversation-design issues, requiring separate treatment.
Why Agents Need More Than Transcription
A transcription service can correct output later. A voice agent must respond within 200 to 300 milliseconds after a caller's turn ends. There is no later correction. A wrong intent produces the wrong next question, and the caller may not be able to correct it before the conversation turns.
Function calls make the impact sharper. A recognised intent can trigger a CRM lookup, product database query, or appointment availability check. If recognition is wrong, the wrong query runs and the reply uses bad data. The error reaches both the next sentence and the system-of-record update after the call.
Accurate accent recognition is table stakes for the Indian market, not a differentiator. Without it, we would not have had early-access teams running production campaigns on Ringg. This phonetic work is what "built for Indian call teams" means technically, and why the first several months went to a problem that can look like narrow infrastructure from outside.