Miso TTS Self-Hosted vs API: Cost, Latency and VRAM Reality

Miso TTS is a credible open-weight option for expressive English voice applications, but self-hosting is not the automatic winner. The model is large, the public 110 ms figure describes Miso's hosted API on H100-class hardware, and the downloadable inference path needs production engineering around it. Most early products should validate demand with the API. Teams with stable volume, strict data boundaries, or on-premises requirements should benchmark self-hosting on their own hardware.

You may have seen the model called Miso One. Miso Labs' official repository and model card call it Miso TTS 8B. This guide uses the official name while answering the buying question behind both searches: should you pay by generated minute, operate the model yourself, or test a hybrid?

Miso TTS 8B facts that change the deployment decision

The useful facts are less viral than the demo clips. They determine whether the model fits your product and infrastructure.

The 110 ms claim is API TTFB, not local end-to-end latency

Miso publishes 110 ms as its hosted service latency. The project's own GitHub documentation adds the boundary that matters: public references to 110 ms describe hosted production API time-to-first-byte on H100-class hardware, not the unoptimized local inference path. It explicitly tells local users to expect materially slower startup and generation on consumer or workstation GPUs.

Time-to-first-byte is also only one clock in a voice agent. The user experiences capture, speech recognition, end-of-turn detection, LLM generation, TTS queueing, first audio, playback buffering and interruption handling. Our Nemotron 3.5 ASR production review shows why an 80 ms audio chunk is not an 80 ms transcript. The same discipline applies here: a 110 ms first byte is not a 110 ms conversation.

Benchmark at least P50 and P95 for request-to-first-audio, real-time factor, full-utterance completion and interruption recovery. Record cold and warm starts separately. A fast median with long cold starts can still be a poor production experience.

What makes the speech expressive?

Miso's architecture is inspired by Sesame CSM. A 7.7B transformer backbone processes interleaved text and audio context and predicts the first Mimi audio codebook index. A smaller 300M decoder predicts the remaining indices across 32 residual vector quantization codebooks. In plain language, the model can represent a much richer range of sound than a single flat audio-token vocabulary and can condition its next voice turn on prior audio, not just text.

That explains the expressive pacing and tone in Miso's published samples. It does not independently prove that the model beats every hosted TTS system on your accents, domain names, numbers or noisy phone audio. Miso's architecture announcement and audio samples are first-party evidence. Treat them as a reason to test, not as a substitute for a blind listening evaluation.

Miso TTS API vs self-hosting

The API is not inherently less serious and self-hosting is not inherently more private. Privacy depends on contracts, retention, logging and architecture. Self-hosting gives you control of the runtime boundary, then makes you responsible for every copy of prompt audio, cloned voice assets, logs, backups and generated files.

What does the Miso API cost?

As of 26 July 2026, Miso's published monthly pricing lists Starter at $20 for 150 included minutes plus $0.15 per extra minute, and Scale at $100 for 1,000 included minutes plus $0.10 per extra minute. Enterprise lists volume pricing below $0.08 per minute and on-premises deployment as a custom option. Prices and plan details can change, so recalculate from the live pricing page before procurement.

Do not compare that per-minute price with only the electricity consumed by a local GPU. Use total monthly cost:

self-hosted TCO = GPU or server + idle headroom + storage and egress + observability + redundancy + security + engineering and on-call

API TCO = plan + generated-minute overage + integration + vendor-risk controls

The break-even question is:

break-even minutes = monthly self-hosted TCO / effective API cost per generated minute

Use utilization, not wishful capacity. A GPU reserved for low and bursty traffic can be more expensive than an API even when its marginal generation cost looks tiny. Conversely, a fully utilized existing GPU fleet can make self-hosting attractive earlier. Our broader local model versus API break-even framework covers the same utilization trap.

Is the model really free and open source?

The code and weights are publicly downloadable and there is no per-minute model license fee for ordinary self-hosting. The precise label is still important. Miso TTS uses a modified MIT license. It grants broad rights to use, modify and sell the software. If a product using it exceeds 50 million monthly active users or $10 million in monthly revenue, the license requires prominent “Miso Labs” attribution in the product interface.

That condition will not affect most teams, but “MIT” and “modified MIT” are not interchangeable in a due-diligence checklist. Record the model version, license text and dependency licenses when you approve the stack. This article is a technical procurement guide, not legal advice.

Miso TTS is a voice component, not a complete voice agent

The released model generates speech. It does not supply production speech recognition, turn detection, interruption handling, dialogue policy, tool execution, abuse prevention or telephony. Miso's own research note says the current system models individual turns and half-duplex audio, while turn-taking and full-duplex conversation remain future work.

A production voice agent still needs:

• streaming ASR and end-of-turn detection;
• an LLM or dialogue system with tool and policy controls;
• TTS serving with queues, cancellation and fallback voices;
• barge-in and interruption behavior;
• consent, retention and voice-asset access controls;
• quality, latency, abuse and cost monitoring.

If the product must speak more than English, Miso TTS 8B cannot be the only voice path today. Route languages to tested alternatives or postpone the model decision until multilingual quality is proven.

Voice cloning changes the risk model

The repository demonstrates prompted generation by passing audio and its transcript as context. That is technically useful and operationally sensitive. A ten-second recording is not just another media file when it can drive a reusable synthetic identity.

Require explicit permission for every cloned voice, bind voice assets to an owner and allowed purpose, restrict export, log generation, rotate watermark keys, and provide a fast revocation path. Generated audio is watermarked by default in the public code. Keep watermarking enabled unless a documented requirement and risk review justify a change.

For EU deployment, Article 50 transparency duties become applicable from 2 August 2026 for covered systems. The European Commission's transparency guidance and Code of Practice address machine-readable marking of generated audio, disclosure for deepfakes, and informing people when they interact with AI. Scope and exceptions depend on the use case. Use our EU AI Act Article 50 implementation checklist as an engineering starting point and obtain legal review for the final interpretation.

Who should choose each deployment path?

A two-week evaluation that produces a buying decision

1. Freeze the test set. Use consented voices and 100 to 300 real scripts covering names, numbers, abbreviations, emotions, long turns and difficult pronunciation.
2. Run the same corpus through API and local paths. Keep prompts, voice context and output settings identical.
3. Measure the full latency curve. Capture cold and warm P50, P95 and P99 request-to-first-audio, real-time factor and completed-utterance time.
4. Blind-rate quality. Score naturalness, speaker similarity, pronunciation, emotional fit and consistency without telling reviewers which path generated the clip.
5. Load-test failure behavior. Test concurrency, queue growth, cancellation, GPU loss, provider failure and fallback voice quality.
6. Price production, not the demo. Model three realistic volume bands and include idle headroom, redundancy and engineering.
7. Review consent and disclosure. Confirm who may clone which voice, where assets live, how they are revoked and how AI audio is marked.

Define pass and kill thresholds before the test. That keeps a striking demo from overruling the operating model. If you need a bounded commercial format, our paid pilot versus proof-of-concept guide shows how to turn technical uncertainty into an explicit decision.

Sources and claim boundaries

Model size, architecture, English-only support, prompted generation, local requirements, download size, hosted-latency boundary and default watermarking come from the MisoTTS repository. The RVQ design, 7.7B backbone, 300M decoder and half-duplex limitation come from Miso Labs' release note. API prices and features come from Miso's pricing page. License wording comes from the published model license. EU dates and transparency measures come from the European Commission. Facts were checked on 26 July 2026. Wavect did not independently reproduce Miso's audio quality or 110 ms API claim.