OpenAI STT

from deepeval.models import OpenAISTTModel
from deepeval.voice import VoiceConfig

stt_model = OpenAISTTModel(
    model="gpt-4o-transcribe",
    language="en",
)
voice_config = VoiceConfig(
    stt_model=stt_model,
    ...,
)

There are ZERO mandatory and EIGHT optional parameters when creating an OpenAISTTModel:

• [Optional] model: A string specifying the name of the transcription model to use. Defaulted to gpt-4o-transcribe.
• [Optional] api_key: A string specifying your OpenAI API key. Defaults to OPENAI_API_KEY if not passed; raises an error at runtime if neither is set.
• [Optional] base_url: A string specifying a custom endpoint to reach the OpenAI API through. Defaulted to None.
• [Optional] language: A string ISO-639-1 code (e.g. "en") pinning the spoken language, which improves accuracy and latency. Pass "auto" to detect the language per utterance instead. Defaulted to None, which leaves the decision to OpenAI.
• [Optional] cost_per_1m_input_tokens: A float overriding the input token price used for cost accounting. Defaults to the built-in price for the model.
• [Optional] cost_per_1m_output_tokens: A float overriding the output token price used for cost accounting. Defaults to the built-in price for the model.
• [Optional] cost_per_minute: A float overriding the per-minute price used for cost accounting. Only applies to models priced by duration, such as whisper-1.
• [Optional] transcription_kwargs: A dictionary of additional parameters forwarded to OpenAI's audio.transcriptions.create(...) call.

First set your API key:

# .env.local
OPENAI_API_KEY=<your-openai-api-key>
DEEPEVAL_STT_MODEL=whisper-1

OpenAI is the transcription provider when no USE_*_STT flag is set, so USE_OPENAI_STT=1 is only needed to switch back after selecting another provider. deepeval set-stt openai --model whisper-1 writes both for you.

You can also pass the model name in code, which wins over DEEPEVAL_STT_MODEL:

from deepeval.voice import VoiceConfig

voice_config = VoiceConfig(
    stt_model="whisper-1",
    ...,
)

A bare model name is shorthand for OpenAISTTModel(model=...), so you can leave stt_model out entirely and transcription still runs on gpt-4o-transcribe. Pass the model object itself to pin a language.