Bias context and vocabulary

Context biasing steers transcription toward the names, terms, and acronyms you expect, so they come out spelled the way you write them. You send it as a list of entries, each one a keyphrase or a description of the recording, in file and realtime transcription alike.

To fix specific terms, send each one as an entry of context_bias. An entry can hold several words, such as a company name.

import os

from mistralai.client import Mistral

client = Mistral(api_key=os.environ["MISTRAL_API_KEY"])

def transcribe(context_bias=None):
    with open("customer_call.wav", "rb") as audio:
        return client.audio.transcriptions.complete(
            model="voxtral-mini-transcribe-3",
            file={"file_name": "customer_call.wav", "content": audio},
            context_bias=context_bias,
        ).text

print("Without bias:", transcribe())
print("With bias:   ", transcribe(["Northwind", "Quillara", "Alder Creek Design", "Rotterdam"]))

import os

from mistralai.client import Mistral

client = Mistral(api_key=os.environ["MISTRAL_API_KEY"])

def transcribe(context_bias=None):
    with open("customer_call.wav", "rb") as audio:
        return client.audio.transcriptions.complete(
            model="voxtral-mini-transcribe-3",
            file={"file_name": "customer_call.wav", "content": audio},
            context_bias=context_bias,
        ).text

print("Without bias:", transcribe())
print("With bias:   ", transcribe(["Northwind", "Quillara", "Alder Creek Design", "Rotterdam"]))

The example transcribes the same call twice, without and with context. The product name and the company name change:

Without bias: Thanks for calling Northwind Support. [...] I ordered two Quilliera headsets last week. [...] I also need an invoice for my company, Older Creek Design. [...]
With bias:    Thanks for calling Northwind Support. [...] I ordered two Quillara headsets last week, [...] I also need an invoice for my company, Alder Creek Design. [...]

Without bias: Thanks for calling Northwind Support. [...] I ordered two Quilliera headsets last week. [...] I also need an invoice for my company, Older Creek Design. [...]
With bias:    Thanks for calling Northwind Support. [...] I ordered two Quillara headsets last week, [...] I also need an invoice for my company, Alder Creek Design. [...]

An entry can also be a sentence or a paragraph that describes the recording: what it's about, who speaks, and which names come up. Send it the same way as a keyphrase.

curl https://api.mistral.ai/v1/audio/transcriptions \
  -H "Authorization: Bearer $MISTRAL_API_KEY" \
  -F model=voxtral-mini-transcribe-3 \
  -F file=@customer_call.wav \
  -F "context_bias=Support call for Northwind about an order of Quillara headsets shipped from the Rotterdam warehouse and invoiced to Alder Creek Design."

curl https://api.mistral.ai/v1/audio/transcriptions \
  -H "Authorization: Bearer $MISTRAL_API_KEY" \
  -F model=voxtral-mini-transcribe-3 \
  -F file=@customer_call.wav \
  -F "context_bias=Support call for Northwind about an order of Quillara headsets shipped from the Rotterdam warehouse and invoiced to Alder Creek Design."

In a realtime session, send context_bias as a JSON array of strings in session.update, before the first audio chunk.

{
  "type": "session.update",
  "session": {
    "audio_format": { "encoding": "pcm_s16le", "sample_rate": 24000 },
    "context_bias": ["Northwind", "Quillara", "Alder Creek Design", "Rotterdam"]
  }
}

{
  "type": "session.update",
  "session": {
    "audio_format": { "encoding": "pcm_s16le", "sample_rate": 24000 },
    "context_bias": ["Northwind", "Quillara", "Alder Creek Design", "Rotterdam"]
  }
}

The context applies to the whole session. After the first audio chunk, the session rejects any change to its fields with an error event.

Both modes take the same entries, sent in a different way:

Context has no budget of its own: its entries count toward the model's context window, together with the audio. A long recording leaves less room for context, so keep the list to the terms and descriptions that matter.

When the audio and the context together go over the window, the request fails and the API doesn't shorten your list: