I want to use the the gpt-4o-transcribe-diarize or gpt-4o-transcribe model as the STT model for my agent. The model is deployed in Azure. How do i use this model?

Thanks for reaching out, Heidi. I'm an AI assistant and I'm here to help. Give me a moment while I research this topic.

I want to transcribe estonian, russian and english. I am using Python. i expect i need realtime straming for a voice bot to work. diarization i do not need but i saw that the model performed better with Estonian language. Should i also set the output to "text" or is that the default

Thanks — I’m responding to your questions about using an Azure-deployed

gpt-4o-transcribe

STT model in Python for real-time transcription of Estonian, Russian, and English, and whether you need to set output="text". Short answers • Use the LiveKit OpenAI plugin and connect via

openai.STT.with_azure(...)

pointing at your Azure endpoint and key. • Real-time streaming is supported when you run the STT inside an AgentSession; audio frames are streamed and transcriptions are produced in real time. • Specify the

language

parameter for best accuracy (especially for Estonian). If you know the speaker’s language, set it explicitly. • You do NOT need to set

output="text"

— plain text transcription is the default output. What to configure (practical checklist) • Environment: ◦ AZURE_OPENAI_API_KEY — your Azure OpenAI key ◦ AZURE_OPENAI_ENDPOINT — e.g. https://<your-resource>.openai.azure.com ◦ OPENAI_API_VERSION — e.g. 2024-08-01-preview (match a version that supports transcribe models) • Model / deployment: ◦ Use

model="gpt-4o-transcribe"

for standard transcription. ◦ You can use

gpt-4o-transcribe-diarize

only if you require diarization (you said you don’t), and diarization must be enabled in your Azure deployment. ◦ If your Azure deployment uses a custom deployment name, pass that deployment name via the plugin option (e.g.,

azure_deployment

). • Language: ◦ For best Estonian accuracy, set

language="et"

. ◦ Use

language="ru"

for Russian and

language="en"

for English. ◦ If you must handle mixed-language sessions dynamically, route audio to the appropriate STT instance or language-aware flow — the plugin expects a language code for best results. Python example

from livekit.plugins import openai
from livekit.agents import AgentSession

# Example: Estonian transcription
stt_et = openai.STT.with_azure(
    model="gpt-4o-transcribe",
    azure_endpoint="https://<your-resource>.<http://openai.azure.com|openai.azure.com>",
    api_key="<YOUR_AZURE_OPENAI_API_KEY>",
    api_version="2024-08-01-preview",
    language="et",
    # azure_deployment="your_deployment_name"  # include if deployment name != model name
)

# Create session (real-time streaming is handled when audio frames arrive)
session = AgentSession(
    stt=stt_et,
    # configure llm, tts, etc.
)

Notes about multilingual realtime usage • If you know each user’s language beforehand, initialize the STT with that language for each user/session. • If you expect frequent language switches within a single stream, you’ll need a strategy to detect language (a lightweight detection step) and switch routing to the correct STT instance. The plugin requires a language code for best results and does not auto-detect reliably for multilingual streams. • For your voice bot latency needs, real-time streaming with AgentSession will provide interim/final transcripts as audio is streamed. About output format • The STT model returns plain text transcripts by default. There’s no

output="text"

parameter you must set for basic transcription — text is the standard output. If you want, next I can: • Provide a ready-made Python example showing how to switch STT instances per participant language in a LiveKit session. • Show an example that handles interim vs final transcripts and integrates them into a voice-bot turn flow. • Help verify the exact

azure_deployment

name and required environment variables for your setup. Which of those would you like next? Sources: Azure OpenAI STT plugin guide | LiveKit Documentation