Work through these in order
1
Confirm what was actually heard
Open the call in Call history and read the transcript. If the transcript is already wrong, the transcriber is the problem. If the transcript is right and the agent still got it wrong, it is a prompt problem — go to Write better prompts.
2
Boost the words that matter
Add brand names, product names and other unusual terms to the Keywords field in the format
word:boost_value — for example Bruce:100. Keyword boosting only takes effect with Deepgram transcription; the field is ignored by other providers.3
Match the transcriber to the language
Providers differ sharply by language. Sarvam is tuned for Indian languages such as Hindi, Tamil and Telugu; Deepgram is strong on low-latency English. On a multilingual agent you can set a different transcription provider per language — see Per-language configuration.
4
Turn on precise transcripts
Enable Generate precise transcript on the Engine tab for higher-accuracy transcription. Use it when accuracy matters more than the small processing cost — see Capture precise transcripts.
5
Cut the noise reaching the transcriber
Enable noise cancellation on the Call tab. Callers on speakerphone in a shop or a car are the most common source of garbled transcripts.
Digits, codes and spellings
Numbers are where mishearing costs the most. Two defenses:- Ask the agent to confirm. Instruct it to read digits back before acting on them — “I have 9 8 7 6, is that right?”. Prompt-level confirmation catches what no transcriber setting will.
- Give the caller a keypad option. For account numbers, OTPs and menu choices, collect keypad input (DTMF) instead of relying on speech.
Code-switching
Indian callers routinely mix English words into Hindi sentences. That is normal input, not an error, and it does not by itself trigger a language switch — see How language switching works for what does. If the agent switches language when it should not, the fix is on that page, not in transcription settings.Next steps
Choose voice and language
Providers, models, keywords
Set up a multilingual agent
Per-language providers
Collect keypad input
Skip speech for digits