Audio Best Practices

Course updated November 3, 2026 for Voxtral TTS 3, based on the Migrate from Voxtral 2 guide and Mistral’s Text to Speech pages: voxtral-mini-tts-2603, deprecated October 31, 2026 and retired April 30, 2027, is replaced in examples by voxtral-mini-tts-3-0. Voxtral 2 predefined voices (gb_jane_neutral…) remain supported; Voxtral 3 adds [emotion=…] and [pace=…] tags in text, the expressiveness setting, custom pronunciation (phoneme_hints), and 35 languages. Mistral recommends re-measuring latency after migration. Sources: Migrate from Voxtral 2 · Speech generation.

Input Quality Determines Output Quality

Voxtral’s voice cloning is powerful, but it remains entirely dependent on the quality of the audio sample you provide. A poor recording will produce a poor synthetic voice, without exception and with no way to fix it later. The following four rules can each be stated in a single sentence, yet they are what separate a voice people enjoy listening to from one they mute after thirty seconds.

Rule 1: Clean Audio

The recording environment is critical because the model cannot distinguish your voice from its surroundings: it learns everything. Constant background noise—ventilation, open-plan office chatter, traffic behind a window—will be captured and reproduced in every generation, even six months later on completely unrelated text. Reverberation causes the same issue: an empty room with hard surfaces, tile, or large windows creates an echo that the model will faithfully reproduce; prefer a furnished room or minimal acoustic treatment. Even very faint background music will be picked up and mixed with the voice. Finally, the sample must contain only one speaker: two overlapping voices yield unusable results.

No studio? That’s not a dealbreaker. A closet full of clothes makes an excellent acoustic isolator and does the job of expensive absorption panels. A good-quality podcast-style headset microphone is more than enough—the gap between that and a professional mic is far smaller than the gap between a quiet room and a noisy one. Record with windows closed and notifications turned off, and make it a habit to run a 5-second test that you listen to before the final recording: that’s when you discover the vibrating fridge or the keyboard sitting too close to the mic.

Rule 2: Speak Naturally

The model captures not only your timbre but also your speaking style, and this is where many first attempts fail. Placed in front of a mic, people spontaneously start articulating like a news anchor—and the synthetic voice inherits that affected tone, which sounds fake by the second sentence. Speak as you normally do, without forcing your diction or altering your usual pace.

The reverse is equally true: a flat, monotone recording will produce a monotone voice. Vary your intonation naturally and keep a steady rhythm—not too fast, or the model misses nuances, and not too slow, or you’ll end up with a sleepy-sounding voice. Keep in mind that the emotion in your sample directly influences the emotion of the generated voice: if you record with enthusiasm, the synthetic voice will sound enthusiastic; if you record calmly, it will sound calm. So choose your recording mood based on the intended use—an audiobook narrator and a support agent don’t call for the same energy.

Rule 3: Avoid Filler Words

Filler words, those placeholders that punctuate spontaneous speech, are the enemy of voice cloning. Avoid parasitic “ums,” “uhs,” “wells,” and “sos,” and don’t cough or clear your throat during recording. If you make a mistake, start over from the beginning rather than correcting mid-stream—a “sorry, let me restart” will remain in the sample.

The reason is simple: the model cannot tell the difference between your voice and your hesitations. It treats them as a characteristic of your speaking style and reproduces them in every generation. A customer who hears their voice assistant systematically hesitate mid-sentence won’t think it’s a bug; they’ll think the service is sloppy.

Rule 4: Optimal Duration

As covered in the previous lesson, the recommended duration is 10 to 20 seconds, but that range deserves nuance. Below 5 seconds, the model doesn’t have enough data to capture your voice variations, and the result sounds generic. Voxtral 2 rejected samples longer than 30 seconds; Voxtral TTS 3 accepts up to a few minutes, but beyond twenty seconds or so, it’s the variety of the speech rather than its length that improves the result. The sweet spot is therefore around 15 to 20 seconds of varied, natural speech—variety counts just as much as length.

The Cross-Lingual Trick

Here’s an advanced technique discovered by Voxtral users: cross-lingual cloning lets you create interesting accent effects. The principle is to provide a sample in one language, say French, and then generate speech in another, say English. The result? The synthetic voice speaks English with the original speaker’s French accent.

# Create a voice from a French sample
voice = client.audio.voices.create(
    name="fr-accent-voice",
    sample_audio=french_sample_b64,
    sample_filename="sample_fr.mp3",
    languages=["fr"],
    gender="female",
)

# Generate speech in English with this voice
response = client.audio.speech.complete(
    model="voxtral-mini-tts-3-0",
    input="Hello, have you tried that new coffee place on Fifth Street?",
    voice_id=voice.id,
    response_format="mp3",
)

The result sounds exactly like the French speaker speaking English, with their natural accent. The most obvious application is for brands that want to maintain their vocal identity across languages: a French company spokesperson keeps their sonic signature on the English version of a campaign instead of being replaced by an anonymous voice.

Pre-Recording Checklist

Before you start your final recording, verify:

• Quiet environment (no background noise, no echo)
• Microphone positioned correctly (10–15 cm from mouth)
• Notifications turned off (phone, computer)
• Script prepared or topic in mind
• 5-second test validated

Key Takeaways

• Clean audio with no background noise or reverberation is essential
• Speak naturally and expressively—the model captures your style
• Avoid filler words (“um,” “uh”) as they will be reproduced
• Optimal duration is 15 to 20 seconds of varied speech
• Cross-lingual cloning lets you preserve an accent in another language