Expressive voice-agent research
Prototype conversational agents where pauses, emphasis, emotion, and turn context matter more than simple narration.
Independent tool overview
Miso One at a glance
Miso One is the launch name associated with Miso Labs' current open release, Miso TTS 8B. It is an English text-to-speech model for expressive conversational speech, dialogue, and voice continuation from optional prompt audio. The weights and local inference code are public under a modified MIT license, but Miso Labs has not launched or priced its promised production API. Running it yourself requires substantial hardware: the official repository recommends at least 24 GB of GPU memory for half-precision inference.
Visit the official Miso One site ↗
- Current model
- Miso TTS 8B
- Best for
- Developers prototyping expressive English voice systems
- Languages
- English only
- Model size
- About 8.2B parameters
- Recommended GPU
- 24 GB VRAM for bf16/fp16
- Hosted API
- Coming soon; no public price
- License
- Modified MIT
Overview
What Miso One is
Miso TTS 8B turns text and optional audio context into conversational speech. Its architecture follows Sesame-style conversational speech modeling: a large Llama-style backbone predicts the first Mimi audio codebook and a smaller autoregressive decoder fills in the remaining codebooks. That design is intended to preserve timing, tone, and dialogue context rather than produce a flat sentence in isolation.
The practical product today is the downloadable model, inference repository, and Miso Labs demo. Developers can run the model locally, condition it with a transcript and a short reference recording, or use earlier generated speech as context for a continued exchange. This makes Miso most relevant to technical teams prototyping expressive agents, characters, and private voice systems—not buyers looking for a polished no-code voice platform.
The headline tradeoff is infrastructure. Miso is unusually open for an 8B-parameter voice model, but the official setup downloads roughly 30–40 GB and recommends a 24 GB GPU. Miso's advertised 110 ms figure refers to time to first audio on H100-class hosted infrastructure, not ordinary local hardware or complete response latency. There is no public API price, service-level commitment, or production usage allowance to compare yet.
Use cases
Who Miso One is best for
The strongest fit depends on the job you need the product to complete, not the size of its feature list.
Local and on-premises deployment
Keep model execution and sensitive voice data on infrastructure you control, subject to the model license and your own security practices.
Voice continuation and cloning experiments
Condition generation on transcript-plus-audio context or a short reference clip to continue a voice and conversational style.
Open-model customization
Inspect the inference code and model architecture instead of relying exclusively on a closed hosted voice API.
Well-equipped ML teams
Use the model when you already have a 24 GB or larger GPU, audio engineering expertise, and the capacity to build a production layer around raw inference.
Capabilities
Core Miso One features
Expressive conversational speech
Generates English speech with dialogue context, timing, and vocal character rather than treating each line as an unrelated narration clip.
Optional audio context
Accepts prompt audio with a matching transcript for voice continuation and one-shot voice-cloning workflows.
Two-stage audio generation
Combines an 8B Llama-style backbone with a roughly 300M-parameter audio decoder to predict 32 Mimi audio codebooks.
Local inference
Public code and model weights let technical teams run the system on their own GPU or private infrastructure.
Dialogue conditioning
Context segments can identify speakers and include prior text and audio so a later response follows the preceding exchange.
Watermarked output
The reference pipeline applies Sony SilentCipher watermarking by default; production deployments should configure a private watermark key.
Public demo
Miso Labs provides a browser demo for evaluating the model without first assembling a local environment.
Process
How the Miso One workflow works
Step 1
Evaluate the official demo
Test representative English scripts, emotional directions, names, punctuation, and dialogue turns before committing infrastructure.
Step 2
Confirm license fit
Review the modified MIT terms, including the Miso Labs attribution requirement for exceptionally large commercial products, with counsel when appropriate.
Step 3
Provision suitable hardware
Plan for roughly 30–40 GB of initial downloads and at least 24 GB of VRAM for the recommended half-precision setup; CPU inference is possible but slow.
Step 4
Install from the official repository
Use the published environment and inference instructions, pin model and dependency versions, and isolate the service from public uploads during evaluation.
Step 5
Add consented reference audio
Only clone voices when you have permission, retain provenance for every reference clip, and block impersonation or deceptive use.
Step 6
Benchmark the complete pipeline
Measure time to first audio, real-time factor, total response delay, GPU utilization, concurrency, quality, and failure rates on your own hardware.
Step 7
Build production controls
Add authentication, quotas, moderation, watermark verification, logging, monitoring, retries, data retention rules, and human escalation around the raw model.
Cost
Miso One pricing and free plan
Miso TTS 8B model weights and inference code can be downloaded without a purchase price under a modified MIT license, but self-hosting compute, storage, engineering, and operations are separate costs. The public demo is available for evaluation. Miso Labs still describes API access as coming soon and has not published usage pricing; enterprise on-premises hosting and support are available by request.
Open model
Free download
Public Miso TTS 8B weights and inference code for local use under the published modified MIT license.
- Self-hosting costs are not included
- Recommended 24 GB GPU for bf16/fp16
- Approximately 30–40 GB first-run download
- Very large commercial products have an additional Miso Labs UI attribution condition
Public demo
Free to try
Browser-based evaluation of Miso's speech output without setting up the model locally.
- Intended for evaluation
- No published production allowance or SLA
- Availability and limits can change
- Do not treat a demo as a production API
Hosted API
Coming soon
Miso Labs has announced future API access but has not published a launch date, rates, quotas, or service levels.
- No public per-character or per-minute price
- No public concurrency limits
- No public uptime commitment
- The advertised hosted latency is not a price or availability guarantee
Enterprise on-premises
Custom
Hosting assistance and support contracts for organizations deploying the model in their own environment.
- Available on request
- Confirm support scope and response times
- Confirm security, data handling, and update terms
- Infrastructure remains a material part of total cost
Pricing checked . Check current pricing at the source ↗
Assessment
Miso One strengths and limitations
Where it stands out
- Public weights and inference code provide substantially more deployment control than a closed speech API.
- Optional transcript-and-audio context supports voice continuation and multi-turn conversational conditioning.
- The model targets emotional, dialogue-oriented speech rather than narration alone.
- Local and on-premises operation can keep sensitive audio inside infrastructure the buyer controls.
- Watermarking is enabled in the reference inference pipeline.
- The official repository documents memory, storage, precision, and latency expectations more candidly than many model launches.
- A public demo lets teams hear the model before provisioning a large GPU.
What to consider
- English is the only language listed by the official repository.
- The recommended 24 GB of GPU memory places local use beyond most consumer laptops and smaller cloud instances.
- Initial model and watermark downloads total roughly 30–40 GB, and CPU inference is described as slow.
- Miso's 110 ms claim is hosted time to first audio on H100-class hardware, not ordinary local performance or end-to-end response time.
- The promised production API still has no public launch date, price, quota, concurrency limit, or service-level commitment.
- The repository is an inference implementation, not a complete production voice-agent platform with telephony, routing, analytics, or workflow tools.
- The license is not standard MIT: products above stated monthly-active-user or monthly-revenue thresholds must display Miso Labs attribution in their interface.
- Voice cloning creates consent, impersonation, fraud, and rights-management risks that the deploying team must control.
- Watermarking helps with provenance but does not replace identity verification, abuse prevention, access controls, or disclosure.
- Vendor performance and quality claims should be validated against your scripts, speakers, accents, hardware, and concurrency requirements.
Compare
Miso One alternatives
The right alternative depends on the specific output, workflow, controls and budget your project requires.
Business Operations
PlayAI
Choose PlayAI when a managed conversational voice platform and production APIs matter more than running open weights yourself.
Explore PlayAI →Content Creator
Seed Audio 1.0
Consider Seed Audio 1.0 when you need a broader audio creation model that can generate speech, music, ambience, and effects in one scene.
Explore Seed Audio 1.0 →Content Creator
ElevenLabs Text To Speech GPT
Use the ElevenLabs Text to Speech GPT for a simpler hosted workflow inside ChatGPT rather than managing an 8B model locally.
Explore ElevenLabs Text To Speech GPT →Questions
Miso One FAQs
What is Miso One?
Miso One is the launch name used for Miso Labs' expressive voice model. The current official open release is labeled Miso TTS 8B.
Is Miso TTS 8B free?
The weights and inference code are free to download under a modified MIT license. GPU compute, storage, engineering, hosting, and operations are not included.
What hardware does Miso TTS need?
Miso Labs recommends at least 24 GB of VRAM for bf16 or fp16 inference. The model can run on a CPU with substantial RAM, but the official documentation warns that it will be slow.
Can Miso clone a voice?
Yes. Miso accepts optional transcript-and-audio context and promotes one-shot cloning from a short reference recording. Only use a voice with the speaker's informed permission.
Does Miso support languages other than English?
No multilingual support is documented in the current official repository; it identifies the model language as English.
Does Miso have an API?
Not as a generally available, publicly priced product. Miso Labs says API access is coming soon, while the current usable options are the demo, local model, or a custom enterprise arrangement.
Is Miso really 110 milliseconds?
Miso Labs reports roughly 110 ms time to first audio for its hosted production setup on H100-class hardware. That is not full response latency and should not be assumed for local GPUs.
Can a company use Miso commercially?
The modified MIT license permits broad commercial use, but products above 50 million monthly active users or $10 million in monthly revenue must prominently display Miso Labs in the user interface. Obtain legal advice for your use case.
Bottom line
Our Miso One verdict
Miso TTS 8B is compelling for experienced teams that want expressive English speech, voice conditioning, and the freedom to run an open model locally. It is not yet a turnkey replacement for a mature hosted voice platform: hardware demands are high, the API remains unpriced and unreleased, and buyers must supply the production, safety, and consent layers. Start with the demo, benchmark on the exact deployment hardware, and choose Miso primarily when model control and on-premises operation justify the engineering cost.
Visit Miso One website ↗Similar Miscellaneous Tools
View all Miscellaneous AI tools →
AutoScientist
Autoscientist - Adaption's new tool for automating AI model training
Gemini 3.5 Live Translate
Gemini 3.5 Live Translate - Google's real-time voice model for live translation across 70+ languages
GPT-Realtime-2
GPT-Realtime-2 - OpenAI's new voice model that can think, call tools, and recover from interruptions in live calls
DiffusionGemma
DiffusionGemma - Google's open diffusion model that can quadruple text generation speed
AI training for the future of work.
Get access to all our AI courses, hundreds of real-world AI use cases, live expert-led workshops, an exclusive network of AI early adopters, and more.
AI Courses
Get unlimited access to all of our current & upcoming industry-specific AI courses for the duration of your subscription.
Daily Guides
To keep up with the rapid pace of AI, our team publishes AI implementation guides daily. Our library contains 300+ practical use cases to automate real-world work.
Workshops
Join weekly, live, interactive sessions with industry leaders who are at the forefront of AI for hands-on implementation guidance and exclusive insights.
Community
Network with an exclusive community of AI-first professionals who are working smarter with AI. Learn how early adopters are using AI in their work and businesses.