Uploaded August 2026 | Updated September 2026, 3 weeks ago
I ripped out my entire voice AI stack, every subscription and API key, and replaced all of it with one thing: Mistral's Voxtral. This is part 2 of my Mistral experiment. Everything runs live, nothing is faked, and if it fails, you watch it fail.
I put it through three rounds on my real production workload: batch transcription across languages, multilingual voice generation, and a live voice agent I can talk to in any language. The code-switching test (English to Hindi to French mid-sentence) is the one that surprised me.
Try Mistral Speech yourself: https://mistr.al/audio-sirajraval-yt
Mistral Speech is a full audio intelligence stack that passes the human test. Voxtral Mini Transcribe 2, Voxtral Realtime, and Voxtral TTS are open-weights models for batch transcription, live STT, and natural voice synthesis. Multilingual, cost-efficient, and deployable on-prem or at the edge.
Follow Mistral on X: https://x.com/mistralai
Subscribe: @MistralAIOfficial
Follow on LinkedIn: linkedin.com/company/mistralai
Chapters:
0:00 My whole business runs on voice
0:33 The rules: everything live, nothing faked
0:47 What Voxtral actually is
1:26 Round 1: transcription (English → Hindi → French, one take)
2:30 The real pricing on my actual files
2:35 Round 2: wiring it into my multi-agent system
2:54 Round 3: live voice agent, real round trip
3:13 My straight verdict: who should switch, who should wait
3:41 Run the same experiment yourself
3:51 Part 1: the agent that coded my whole codebase
What would you build with a multilingual voice agent? Tell me in the comments.
🔔 Subscribe for more real-world AI builds: youtube.com/@SirajRaval
📬 Business inquiries: hello@sirajraval.com
📲 Follow
X: https://x.com/sirajraval
Instagram: instagram.com/sirajraval
LinkedIn: linkedin.com/in/sirajraval
#AI #Mistral #Voxtral #VoiceAI
I ripped out my entire voice AI stack, every subscription and API key, and replaced all of it with one thing: Mistral's Voxtral. This is part 2 of my Mistral experiment. Everything runs live, nothing is faked, and if it fails, you watch it fail.
I put it through three rounds on my real production workload: batch transcription across languages, multilingual voice generation, and a live voice agent I can talk to in any language. The code-switching test (English to Hindi to French mid-sentence) is the one that surprised me.
Try Mistral Speech yourself: https://mistr.al/audio-sirajraval-yt
Mistral Speech is a full audio intelligence stack that passes the human test. Voxtral Mini Transcribe 2, Voxtral Realtime, and Voxtral TTS are open-weights models for batch transcription, live STT, and natural voice synthesis. Multilingual, cost-efficient, and deployable on-prem or at the edge.
Follow Mistral on X: https://x.com/mistralai
Subscribe: @MistralAIOfficial
Follow on LinkedIn: linkedin.com/company/mistralai
Chapters:
0:00 My whole business runs on voice
0:33 The rules: everything live, nothing faked
0:47 What Voxtral actually is
1:26 Round 1: transcription (English → Hindi → French, one take)
2:30 The real pricing on my actual files
2:35 Round 2: wiring it into my multi-agent system
2:54 Round 3: live voice agent, real round trip
3:13 My straight verdict: who should switch, who should wait
3:41 Run the same experiment yourself
3:51 Part 1: the agent that coded my whole codebase
What would you build with a multilingual voice agent? Tell me in the comments.
🔔 Subscribe for more real-world AI builds: youtube.com/@SirajRaval
📬 Business inquiries: hello@sirajraval.com
📲 Follow
X: https://x.com/sirajraval
Instagram: instagram.com/sirajraval
LinkedIn: linkedin.com/in/sirajraval
#AI #Mistral #Voxtral #VoiceAI
