Fish Audio / OpenAudio S1 + S2 vs Voxtral

June 5, 2026 update: this comparison has been rewritten because Voxtral is no longer only an STT buyer conversation. Mistral now has Voxtral TTS and STT in the same ecosystem.

Fish Audio and Voxtral but still come from different buying logic. Fish Audio is the better first test when you want OpenAudio S1/S2 control, open weights, batch transcription, and realtime STT under one provider.

Quick Answer

, self-hosting, voice-cloning experiments, or API on the output side depending on quality, latency, and governance tests.

Decision Snapshot

• Primary job: Fish Audio / OpenAudio S1 + S2 and voice cloning, transcription, and realtime STT.
• Flagship audio pieces: Fish Audio / OpenAudio S1 + S2: OpenAudio S1, OpenAudio S2/S2 Pro, hosted Fish Audio API. Voxtral: Voxtral TTS v26.03, Voxtral Mini Transcribe 2, Voxtral Realtime.
• Pricing shape: Fish Audio / OpenAudio S1 + S2: Free/open-weight path plus hosted S1/S2 Pro at $15 per 1M UTF-8 bytes; creator tiers from free to Max $749/mo. Voxtral: Voxtral TTS at $0.016 per 1K characters; Mistral lists Voxtral Mini Transcribe at $0.002/min while the Mini Transcribe 2 card shows $0.003/min.
• Best for: Fish Audio / OpenAudio S1 + S2: Custom speech generation, local deployment, open-source model evaluation, agent spoken output. Voxtral: Mistral-standardized apps, speech input loops, multilingual ASR, and hosted TTS/STT procurement.
• Main risk: Fish Audio / OpenAudio S1 + S2: More operational QA and safety policy owned by your team. Voxtral: Less control outside Mistral’s platform and model/license boundaries.

Where Fish Audio Wins

• Better when open weights, MIT licensing, and self-hosting are procurement requirements.
• Stronger for developers who want to inspect, tune, and operate the TTS layer rather than only call a hosted API.
• More natural fit for custom voices, narration systems, agent spoken replies, and model-quality experiments.
• API pricing is easy to reason about for S1/S2 Pro: $15 per 1M UTF-8 bytes, with Fish estimating roughly 180,000 English words or about 12 hours of speech per 1M UTF-8 bytes.
• Better when you need a fallback path if hosted provider pricing, rate limits, or data policies change.

Where Voxtral Wins

• Better when your product already uses Mistral for LLMs and wants audio procurement, keys, monitoring, and vendor review in the same ecosystem.
• , 9-language support, streaming, and around 90ms time-to-first-audio in the model card.
• Voxtral Mini Transcribe 2 covers batch STT with diarization, context biasing, word-level timestamps, 13-language support, and up to 3-hour recordings.
• Voxtral Realtime is built for live transcription with configurable sub-200ms latency and open weights under Apache 2.0.
• Better fit when the buyer question is the whole voice loop, not just speech output.

Key Differences

, batch transcription, and realtime transcription can live beside the same LLM platform.

The pricing units are not apples-to-apples. Fish bills hosted S1/S2 Pro by UTF-8 input bytes. Mistral bills Voxtral TTS by characters and bills transcription separately by audio minute/model. For a real budget, run the same scripts through both systems and compare delivered audio minutes, latency, pronunciation fixes, retries, rate limits, and engineering time.

Who should choose Fish Audio

Choose Fish Audio when you need open-weight TTS, custom narration, character voices, voice cloning, agent spoken output, self-hosting, or a technical escape hatch from pure hosted APIs.

Who should choose Voxtral

Choose Voxtral when you need Mistral-native audio: hosted TTS, transcription, live ASR for a voice agent, multilingual audio understanding, or a unified provider story for speech input and output.

Bottom Line

Fish Audio is still the stronger open-control TTS pick. Voxtral is now a serious Mistral-native audio stack rather than only an STT comparison. Pick Fish when you want to own the speech model path; pick Voxtral when staying inside Mistral for TTS plus STT reduces product and procurement friction.