Last updated
What is OpenAI speech-to-text used for in a Bolna voice agent?
OpenAI speech-to-text transcribes the caller for your Bolna agent with gpt-realtime-whisper, a streaming model that returns partial text while the caller is still speaking. Voice activity detection is built in, so it finds the start and end of speech without manual setup. A delay setting lets you trade latency against accuracy for accented speech or noisy lines.
How do I set up OpenAI speech-to-text with Bolna?
- Open your agent at platform.bolna.ai and go to the Languages tab.
- Under Transcription, set Provider to
openaiand Model toGPT Realtime Whisper. - Set the language as an ISO 639-1 code, such as
enorhi. - Choose a Transcription delay. Keep
mediumunless you have a reason to change it. - Turn on Noise reduction for call-center or office audio.
In the API, set provider to openai and model to gpt-realtime-whisper, and tune delay and noise_reduction. See the OpenAI transcriber guide in the Bolna docs.
Which delay settings does OpenAI speech-to-text support on Bolna?
delay value | Behavior |
|---|---|
minimal | Fastest, best for simple, clean audio |
low | Low latency with good accuracy |
medium | Default, balanced for most agents |
high | Higher accuracy for accents or background noise |
xhigh | Maximum accuracy, highest latency |
Common language codes include en, es, fr, de, hi, pt, ja, it, nl, zh, ko, ar and ru.
What can you build with OpenAI speech-to-text and Bolna?
Agents for accented or noisy callers. Raise delay to high and turn on noise reduction so the model hears more audio before it commits to a transcript.
Fast agents on clean lines. Set delay to minimal or low when audio is clear. Interim transcripts give the LLM an early start on its reply.
OpenAI stacks with your choice of voice. Pair the transcriber with an OpenAI LLM and any voice, such as ElevenLabs or Cartesia. For a single model that hears and speaks, see OpenAI Realtime.
Multilingual agents. GPT Realtime Whisper supports a wide range of languages, including Hindi, Japanese, Chinese, Korean and Arabic.