Key Takeaways
- Izwi (github.com/izwi-ai/izwi) is a free, open-source, local-first voice AI runtime for text-to-speech, speech-to-text, and related audio tasks
- "Izwi" is the Zulu and Xhosa word for "voice"
- MIT licensed, confirmed via the GitHub repository's LICENSE file
- Ships as a desktop app (macOS, Linux, Windows), a web UI, and a command-line tool, all built on the same local inference server
- Over 383 GitHub stars and 40 forks as of this review
- Still in beta: the most recent tagged release is v0.1.0-beta-17 (June 22, 2026), even though the repository's most recent commit push is dated September 13, 2026
π In One Sentence
Izwi is a free, open-source (MIT), local-first voice AI runtime β desktop app, web UI, and CLI β that bundles text-to-speech, speech-to-text, speaker diarization, and voice cloning behind one OpenAI-compatible API, with 383+ GitHub stars.
π¬ In Plain Terms
Izwi is software you install on your own computer that turns text into spoken audio, turns spoken audio into text, and can clone a voice from a sample β all running locally instead of through paid cloud APIs like those from ElevenLabs or OpenAI's audio endpoints. Because its API matches OpenAI's own format, many existing apps built for that API can point at Izwi instead with minimal changes.
πNote: This review is based on Izwi's own GitHub repository, its README, and its release history via the GitHub API. It does not claim PromptQuorum has independently benchmarked the audio quality or transcription accuracy of any specific model Izwi supports.
What Is Izwi?
Izwi is a free, open-source, local-first voice AI runtime that bundles several separate voice tasks β text-to-speech, speech-to-text, diarization, and voice cloning β behind one OpenAI-compatible API, instead of requiring a different cloud service for each. Its own README describes it as "local-first voice AI for speech, chat, and audio workflows."
- Product type: a local inference server with three front ends β a native desktop app, a browser-based web UI, and a CLI β all talking to the same underlying engine
- Repository: github.com/izwi-ai/izwi, created January 23, 2026
- License: MIT, confirmed via the repository's LICENSE file
- Website: izwiai.com, with separate documentation at izwiai.com/docs
- Funding: no funding round, investor, or commercial backing was found for this review β treat Izwi as an apparently independent, community-supported open-source project
- Scale: over 383 GitHub stars, 40 forks, as of this review
What Can You Do With Izwi?
Izwi covers four related voice tasks β speaking, listening, cloning, and organizing local models β through one consistent interface, whether you use the desktop app, web UI, or CLI.
- Text-to-speech: convert text into spoken audio, including long-form "Studio" projects, with support for voice design and saved custom voices
- Voice cloning: generate speech in a cloned voice from a reference sample, using supported TTS model families
- Speech-to-text: transcribe audio files, plus realtime streaming transcription for live audio
- Speaker diarization and forced alignment: identify which speaker said what in multi-speaker audio, and align transcribed text to exact audio timestamps
- Real-time voice conversation: combine local automatic speech recognition, a local chat model, and local text-to-speech for a spoken back-and-forth, similar in concept to a voice assistant
- Model management: download, load, unload, and delete models from within the app; view chat history and export results
- OpenAI-compatible API:
/v1routes for models, chat completions, audio speech, and audio transcriptions, plus preview support for the Responses API shape, so existing OpenAI-API client code can often point at a local Izwi server with minimal changes
Which Models Does Izwi Support?
Izwi is a runtime, not a single model β it supports several interchangeable model families per task, so you pick the one that fits your hardware and quality needs.
Text-to-speech
- Supported model families (per Izwi's catalog):
- Qwen3-TTS, Kokoro-82M, Voxtral TTS, VibeVoice
Speech-to-text
- Supported model families (per Izwi's catalog):
- Parakeet, Whisper, Qwen3-ASR, Nemotron 3.5 ASR, VibeVoice ASR, LFM2.5 Audio, Voxtral Mini
Diarization & alignment
- Supported model families (per Izwi's catalog):
- Sortformer (diarization), Qwen3 ForcedAligner (timestamp alignment)
Chat
- Supported model families (per Izwi's catalog):
- Qwen3, Qwen3.5, LFM2.5, Gemma
Run izwi list locally to see the current enabled catalog, since supported models can be added or changed between releases. Some model weights carry their own separate licenses or usage restrictions β check Izwi's Models Guide on izwiai.com/docs/models before redistribution or commercial use.
Usage Examples
Izwi's CLI covers the core tasks in a handful of commands, once a model is downloaded.
# Start the local server with the web UI
izwi serve --mode web
# Download a TTS model, then generate speech
izwi pull Qwen3-TTS-12Hz-0.6B-Base
izwi tts "Hello from Izwi." --output hello.wav
# Download a speech-to-text model, then transcribe a file
izwi pull Parakeet-TDT-0.6B-v3
izwi transcribe audio.wav --model Parakeet-TDT-0.6B-v3Platform, Pricing, and Licensing
Platform
- What Izwi states:
- Desktop app for macOS, Linux, and Windows; also available as a web UI and CLI/server, all built on the same local engine.
Cost
- What Izwi states:
- Free and open source. No account, subscription, or API key is required for local inference; some model weights may carry their own separate license terms.
Licensing
- What Izwi states:
- MIT, confirmed via the GitHub repository's LICENSE file.
Hardware acceleration
- What Izwi states:
- macOS Apple Silicon release builds use Metal on macOS 15+ and fall back to CPU on macOS 12-14. Linux and Windows release builds are CPU-only by default; NVIDIA CUDA is supported through a Docker CUDA profile or a source build.
Status
- What Izwi states:
- Still in beta β its most recent tagged release is v0.1.0-beta-17, with commit activity continuing after that tag.
Verify the current Runtime Support Matrix directly on izwiai.com/docs/support-matrix before installing, since hardware acceleration support can change between releases.
Install Izwi
Izwi ships pre-built installers for all three desktop platforms, plus a CLI install script and source builds.
Source | Link |
|---|---|
| GitHub Releases (macOS .dmg, Linux .deb, Windows .exe) | github.com/izwi-ai/izwi/releases |
| GitHub repository (source code, MIT) | github.com/izwi-ai/izwi |
| CLI install script | ./scripts/install-cli.sh, or build manually with cargo build --release -p izwi-cli |
| Documentation | izwiai.com/docs |
On Linux, install the downloaded package with sudo dpkg -i izwi_*.deb. CUDA acceleration on Linux or Windows requires the Docker CUDA profile or a source build with the matching CUDA toolkit β release builds are CPU-only by default.
Izwi vs. Whisper.cpp + Piper
Izwi replaces a common two-tool local voice stack β Whisper.cpp for transcription plus Piper for text-to-speech β with one runtime and one API, at the cost of being a younger, still-beta project.
Scope
- Izwi:
- One runtime covering TTS, STT, diarization, alignment, and voice cloning behind one API
- Whisper.cpp + Piper (separately):
- Two separate projects β Whisper.cpp for STT only, Piper for TTS only β that you run and integrate independently
API shape
- Izwi:
- OpenAI-compatible
/v1routes, so existing OpenAI-API client code often works with minimal changes - Whisper.cpp + Piper (separately):
- Each tool has its own CLI/library interface; no shared or OpenAI-compatible API across both by default
Voice cloning
- Izwi:
- Supported through compatible TTS model families
- Whisper.cpp + Piper (separately):
- Not supported by Piper or Whisper.cpp themselves
Project maturity
- Izwi:
- Beta releases only (latest: v0.1.0-beta-17); created January 2026
- Whisper.cpp + Piper (separately):
- Both are long-established, widely deployed projects individually
Interfaces
- Izwi:
- Desktop app, web UI, and CLI from one project
- Whisper.cpp + Piper (separately):
- Command-line/library tools; any GUI comes from third-party wrappers
Choose Izwi if you want one runtime and one API surface for several voice tasks, including cloning. Choose Whisper.cpp and Piper separately if you specifically want each tool's long individual track record, or only need one of the two tasks. See the Whisper.cpp review and Piper TTS review for details on each.
Who Should Use Izwi?
Izwi fits developers and technical users who want one local tool covering several voice tasks instead of stitching together separate cloud APIs or single-purpose local tools.
What Izwi Is Not Good For
Izwi is not a good fit if you need out-of-the-box GPU acceleration on Linux/Windows or a long-established, non-beta product.
- Not GPU-accelerated by default on Linux or Windows β release builds run on CPU only unless you use the Docker CUDA profile or build from source
- Not a 1.0 release β the most recent tagged version is a beta (v0.1.0-beta-17), so expect the API and feature set to still be settling
- Not a single-purpose, minimal tool β if you only need one task (say, just TTS), a dedicated tool like Piper may be simpler to run
- Not guaranteed to have a tagged release matching its latest code β this review found a gap between the most recent commit push (September 13, 2026) and the most recent tagged release (June 22, 2026)
- Not backed by a funding round or company that this review could verify β treat it as an independently maintained, community-supported project
Common Mistakes When Evaluating Izwi
Most confusion about Izwi comes from assuming it is a single model, expecting GPU acceleration everywhere by default, or missing that it is still pre-1.0.
Competitors and Alternatives
Izwi is most directly comparable to other local or open-source text-to-speech and speech-to-text tools, several of which it can substitute for behind one unified API.
Tool | Best known for | Link |
|---|---|---|
| Whisper.cpp | Fast, widely used local speech-to-text engine, a C/C++ port of OpenAI's Whisper | Whisper.cpp review |
| Piper | Lightweight, widely deployed local text-to-speech engine, often used on low-power devices | Piper TTS review |
| Coqui TTS | Open-source text-to-speech toolkit with voice cloning support | Coqui TTS review |
| MacWhisper | macOS desktop app for local transcription built on Whisper | MacWhisper review |
This list reflects tools commonly compared to Izwi in the local voice-AI space, not an independent PromptQuorum ranking β verify each tool's current feature set and hardware support before choosing.
Frequently Asked Questions
What is Izwi?
Izwi (github.com/izwi-ai/izwi) is a free, open-source (MIT), local-first voice AI runtime that bundles text-to-speech, speech-to-text, diarization, and voice cloning behind one OpenAI-compatible API.
Is Izwi free?
Yes, Izwi is free and open source (MIT license). No account, subscription, or API key is required for local inference; some downloaded model weights may carry their own separate license terms.
What does "Izwi" mean?
"Izwi" is the Zulu and Xhosa word for "voice."
Does Izwi run on Windows?
Yes, Izwi has a Windows installer (.exe), alongside macOS (.dmg) and Linux (.deb) builds. Windows and Linux release builds are CPU-only by default; GPU acceleration via Metal is available on Apple Silicon Macs.
Does Izwi support voice cloning?
Yes, through compatible text-to-speech model families in its catalog, run entirely on your own machine.
Which speech-to-text models does Izwi support?
Parakeet, Whisper, Qwen3-ASR, Nemotron 3.5 ASR, VibeVoice ASR, LFM2.5 Audio, and Voxtral Mini, per Izwi's current model catalog β run izwi list locally to see what is enabled.
Does Izwi use an OpenAI-compatible API?
Yes. It exposes /v1 routes for models, chat completions, audio speech, and audio transcriptions, plus preview support for the Responses API shape.
Is Izwi a finished 1.0 product?
No. As of this review, the most recent tagged release is a beta (v0.1.0-beta-17), so expect the API and model catalog to keep changing.
Does Izwi send my audio or transcripts anywhere?
Per Izwi's own README, inference data stays on your machine. Optional anonymous desktop analytics are off by default and, if enabled, are stated not to include prompts, transcripts, audio payloads, file paths, or personal identifiers.
How do I install Izwi?
Download the installer for your platform from GitHub Releases (.dmg for macOS, .deb for Linux, .exe for Windows), or install just the CLI/server with the project's install script or a source build.