Skip to main content
Kokoro

Kokoro 82M v1.0

Kokoro 82M v1.0 is a compact text-to-speech model from Kokoro with roughly 82 million parameters, used to synthesize spoken audio from written text.

From
$0.990 / 1M chars
across 1 provider
Compare Prices

API Pricing

ProviderPrice /1M char
Deep Infra logo
Deep Infra
$0.990/1M char

Prices updated daily. Last check: Nov 8, 2026

Model Details

General

Creator
Kokoro
Modalities
Text

Capabilities

Strengths & Limitations

Strengths

  • Small footprint at approximately 82 million parameters, which keeps memory requirements and compute cost per request low relative to larger neural TTS systems
  • Synthesis speed benefits directly from the compact size, making it suitable for interactive or streaming-style voice output
  • Available through multiple inference providers, so pricing and regional availability can be compared rather than being tied to a single vendor
  • Tagged v1.0 release gives a stable reference point for reproducible deployments, as opposed to a rolling preview build
  • Model size is small enough to be practical for self-hosting on commodity GPUs or even CPU-bound setups, reducing dependence on a hosted API
  • Fits high-volume batch narration workloads where per-character cost dominates the decision

Limitations

  • At roughly 82M parameters it is a small model, and output naturalness will generally not match much larger commercial TTS systems on expressive or emotional speech
  • Text-to-speech only — it does not transcribe audio or handle conversational turn-taking, so it must be paired with a separate speech-to-text model for a full voice loop
  • Voice inventory, language coverage, and output sample rate vary by host and are not fields we track centrally; verify them with your chosen provider
  • Fine-grained prosody, emphasis, and pronunciation control depend on what each provider exposes rather than being guaranteed by the model entry itself
  • Handling of unusual proper nouns, acronyms, and domain jargon typically requires manual text preprocessing

Key Features

•Text-to-speech synthesis producing audio output from written input
•Approximately 82 million parameters (compact architecture)
•Tagged v1.0 release for deployment stability
•Low-latency generation suited to interactive voice output
•Multiple voice options, as exposed by the hosting provider
•Available via hosted inference APIs with multi-provider pricing comparison
•Small enough footprint for self-hosted or edge-adjacent deployment
•Batch-friendly for long-form narration workloads

About Kokoro 82M v1.0

Kokoro 82M v1.0 is a text-to-speech (TTS) model released under the Kokoro name. The "82M" in the model name refers to its approximate parameter count of 82 million, which places it well below the size of most general-purpose speech synthesis stacks. The "v1.0" designation marks it as a tagged release point in the Kokoro line rather than a preview or experimental build, and it is the version most commonly exposed by inference providers that host the model. As a text-to-speech model, Kokoro 82M v1.0 takes written text as input and returns synthesized audio rather than tokens. Because of this, the specifications that matter for it are different from those of a chat model: there is no context window or reasoning benchmark to compare, and the relevant characteristics are voice selection, synthesis latency, audio sample rate, and how naturally the output renders punctuation, pacing, and sentence boundaries. The small parameter count is the model's defining structural trait — it is sized so that generation can run quickly and at low memory cost, including on modest hardware, which is the main reason it appears across hosted inference APIs and self-hosted deployments alike. In practice, Kokoro 82M v1.0 is used where a large volume of text needs to be turned into audio without a per-request budget that supports larger neural voice systems: narration for articles and documentation, voice output for assistants and chatbots, accessibility read-aloud features, and prototyping of voice interfaces. Buyers typically evaluate it against larger hosted TTS offerings on a quality-versus-cost-and-latency basis, and against other small open TTS models on voice variety and deployment footprint. We do not track every provider-specific detail — supported voices, languages, and output formats can differ between the hosts listed in the pricing table on this page, so confirm those against the provider you plan to use.

Common Use Cases

Kokoro 82M v1.0 fits workloads where text needs to be converted to speech at volume and at low cost per character: article and blog narration, audiobook or documentation read-aloud features, accessibility tooling that voices on-screen content, IVR and notification prompts, and spoken responses for chatbots and voice assistants. Its compact size makes it a reasonable choice for interactive applications where synthesis latency is user-visible, and for self-hosted pipelines where keeping audio generation on your own infrastructure matters for data handling or cost control. It is less appropriate for projects that need highly expressive, character-driven performance or precise voice cloning, where larger specialized speech models are typically evaluated instead. Because it only produces audio, applications that need two-way voice will pair it with a separate speech-to-text model.

Frequently Asked Questions

How much does Kokoro 82M v1.0 cost to use?

Pricing depends on the provider and on how the workload is billed — text-to-speech is commonly metered per character or per unit of generated audio, and some hosts bill for GPU time instead if you run the model yourself. Rates also differ between on-demand and committed capacity. Check the pricing table on this page for current figures from each tracked provider.

What is Kokoro 82M v1.0 best used for?

Converting written text into spoken audio at volume — article narration, read-aloud accessibility features, prompts and notifications, and spoken output for assistants. Its small parameter count makes it most attractive where synthesis latency and per-character cost matter more than maximally expressive delivery.

Can Kokoro 82M v1.0 transcribe audio or hold a conversation?

No. It is a text-to-speech model, meaning it only goes from text to audio. A full voice application needs a separate speech-to-text model for transcription and a language model for the conversational logic, with Kokoro handling only the final audio rendering step.

Why does the 82M parameter count matter?

It is the model's main structural differentiator. At roughly 82 million parameters, Kokoro is far smaller than typical large neural speech systems, which reduces memory use, speeds up generation, and lowers the hardware bar for self-hosting. The trade-off is that very small models generally have less headroom for nuanced prosody and expressive performance than much larger ones.

What voices and languages does it support?

Voice inventory and language coverage are exposed by the provider hosting the model, and we do not track those fields per-provider. Check the documentation of the specific host you plan to use, since the available voice list and output audio format can differ between deployments of the same model.