Audio Transcription Use Cases

Course updated November 3, 2026 for Voxtral 3, based on Mistral’s Migrate from Voxtral 2 guide: Voxtral 2 models (voxtral-mini-2602, voxtral-mini-transcribe-realtime-2602, voxtral-mini-tts-2603), deprecated on October 31, 2026 and retired on April 30, 2027, are replaced in examples by voxtral-mini-transcribe-3, voxtral-mini-transcribe-realtime-3, and voxtral-mini-tts-3-0. Key changes: segment-level timestamping only (word returns a 400 error), real-time diarization support, real-time session settings locked after the first audio chunk, and flexible context_bias inputs (phrases with spaces). Sources: Migrate from Voxtral 2 · Speech to Text.

Why automate audio transcription

Automatic transcription converts an audio stream into structured, usable text. With Voxtral’s current performance, the error rate is low enough for direct use in many professional contexts without exhaustive proofreading. This detail changes the economic equation: as long as line-by-line review remains necessary, automation merely shifts the work rather than eliminating it.

The use cases below share a common trait: each relies on a specific API parameter. Learning to recognize them saves you from manually rebuilding what the model already produces.

Automatic subtitling

Subtitling is the most direct application. Voxtral Mini Transcribe 3’s segment-level timestamping provides the start and end time of each transcription segment, which is sufficient to generate SRT or VTT files synchronized with speech. You enable this output with timestamp_granularities=["segment"]; word-level timestamping, available in Voxtral 2, now returns a 400 error. If a segment is too long for the screen, you split it further on the client side.

In practice, an e-learning organization uses this to subtitle its video catalog, a government agency to make content accessible to deaf and hard-of-hearing users, and a publisher to produce multilingual subtitles by chaining transcription and translation. For live webinars or conferences, the need is identical but shifts to Voxtral Realtime, since the file isn’t available when subtitles must appear.

Example: generating SRT subtitles

from mistralai.client import Mistral

client = Mistral(api_key="YOUR_API_KEY")

with open("video_audio.mp3", "rb") as f:
    response = client.audio.transcriptions.complete(
        model="voxtral-mini-transcribe-3",
        file={"content": f, "file_name": "video_audio.mp3"},
        timestamp_granularities=["segment"]
    )

# Generate an SRT file: one subtitle per segment
def to_srt(segments):
    srt_lines = []
    for idx, seg in enumerate(segments, 1):
        srt_lines.append(f"{idx}")
        srt_lines.append(f"{format_time(seg.start)} --> {format_time(seg.end)}")
        srt_lines.append(seg.text.strip())
        srt_lines.append("")
    return "\n".join(srt_lines)

def format_time(seconds):
    h = int(seconds // 3600)
    m = int((seconds % 3600) // 60)
    s = int(seconds % 60)
    ms = int((seconds % 1) * 1000)
    return f"{h:02d}:{m:02d}:{s:02d},{ms:03d}"

Meeting minutes

Combining transcription + diarization makes Voxtral a serious tool for automatic meeting minutes. The model identifies each speaker and produces a transcription structured by participant, whereas a flat transcription forces the reader to guess who is speaking. Two parameters suffice: diarize=True to separate speakers and timestamp_granularities=["segment"] to obtain time blocks.

A team uses this for weekly check-ins, an executive board for committee meetings, and a quality department to archive and search meeting history. The diarized text also serves as raw material for automatically extracting decisions and action items.

Typical workflow

response = client.audio.transcriptions.complete(
    model="voxtral-mini-transcribe-3",
    file={"content": audio_file, "file_name": "reunion.mp3"},
    diarize=True,
    timestamp_granularities=["segment"]
)

# Each segment contains: speaker, start, end, text
for segment in response.segments:
    print(f"[{segment.speaker_id}] ({segment.start:.1f}s) {segment.text}")

You can then send the result to an LLM (Mistral Large 3, for example) to extract a summary, decisions made, and action items.

Podcasts and interviews

Podcasts and interviews are long-form content—thirty minutes to over two hours—with two to five participants. Voxtral distinguishes the host from guests using diarization, and respects speaker names and subject-specific jargon thanks to context biasing. For the longest recordings, keep the context window in mind, shared between audio and vocabulary: a request exceeding it is rejected with a 400 error (nothing is truncated); split the audio instead.

A podcast producer uses this to generate a full transcription published as an article, full-text indexing for searchable archives, automatically generated show notes, and quotes ready to clip for social media. The same audio file thus feeds four different deliverables.

Digital accessibility

Automatic transcription is a pillar of digital accessibility, mandated by the European Accessibility Act (June 2025). Subtitling video content is now a legal requirement, alongside transcribing voicemails to text, providing text descriptions of audio content—which incidentally benefits SEO—and maintaining accessible archives of recordings.

Voxtral lets you automate these obligations at scale, with a marginal cost per minute of audio far lower than human transcription. For a catalog of several hundred hours, this is the difference between a budgeted compliance rollout and one postponed indefinitely.

Other applications

Before writing any code, identify the use case closest to yours in this list and note the parameters it requires. This combination—not the model alone—determines your result quality.

Key takeaways

• Subtitling uses segment-level timestamping to generate SRT/VTT files
• Meeting minutes combine diarization and segmented transcription
• Podcasts and interviews benefit from long audio support (3h) and context biasing
• Digital accessibility is a legal requirement that Voxtral helps automate
• Each use case relies on specific Voxtral API parameters