What if developers could build voice-enabled AI applications without depending entirely on paid third-party voice APIs?
That is the idea behind RagLeap Voice.
RagLeap Voice is an open-source, MIT-licensed Python project focused on self-hosted voice infrastructure. Its long-term ambition is to provide an open alternative in the space occupied by services such as ElevenLabs and Deepgram.
The goal is simple:
Own your voice infrastructure. Control your data. Avoid vendor lock-in.
GitHub: https://github.com/ragleap/ragleap-core/tree/main/packages/ragleap-voice
PyPI: https://pypi.org/project/ragleap-voice/
Why build RagLeap Voice?
Many voice applications depend on external speech-to-text and text-to-speech providers. This can introduce usage-based costs, paid API-key dependencies, vendor lock-in, and less control over where voice data is processed and stored.
Our direction is to give developers another choice:
- Self-host voice processing.
- Reduce dependence on paid STT/TTS APIs.
- Keep voice data and assets under your control.
- Choose and change engines without permanent vendor dependency.
- Explore training and customizing your own voices.
- Build multilingual voice capabilities.
- Use an open-source MIT-licensed project.
What we're working toward
Self-hosted voice infrastructure
The long-term goal is to run voice processing on infrastructure controlled by the developer or organization.
Voice data ownership
We plan to keep voice data and assets under operator control, outside the primary application database, rather than requiring vendor-hosted storage.
Custom voice training
We want to provide an option for developers and organizations to train or customize their own voices, with a focus on high-quality speech output.
Multilingual capabilities
We want voice processing to support more language communities, not just English.
No vendor lock-in
Developers should have flexibility over models, engines, storage, and deployment.
What is available today?
RagLeap Voice is currently pre-alpha. The latest release is 0.3.1.
Recent development:
| Release | Main additions |
|---|---|
| 0.2.0 | Multilingual cardinal normalization and stateful audio resamplers |
| 0.3.0 | Streaming-oriented building blocks, StreamGate, and Hindi normalization improvements |
| 0.3.1 | Top-level streaming function exports and G.711 A-law helpers |
Import the current streaming functions:
from ragleap_voice import transcribe_stream, synthesize_stream
Important: these interfaces do not yet provide true incremental streaming.
-
transcribe_stream()buffers audio and runs transcription at the end. -
synthesize_stream()chunks playback after the complete waveform has been generated.
Current development status
We want to be transparent about the project's maturity:
- No speech model is verified.
- No model is configured as a default.
- Unverified model calls require
allow_unverified=True. - Real phone-audio performance has not been measured.
- No latency or quality benchmark is claimed.
- Python 3.13 has not been measured for version 0.3.1.
The default path does not require a paid API. Optional engines can be installed separately.
Install
pip install ragleap-voice==0.3.1
For the optional Whisper integration:
pip install "ragleap-voice[whisper]"
PyPI: https://pypi.org/project/ragleap-voice/0.3.1/
The bigger goal
We are not presenting RagLeap Voice as a finished replacement for established voice platforms today.
We are building toward an open alternative: self-hosted voice infrastructure, multilingual capabilities, customizable voices, operator-controlled data, and freedom from permanent vendor dependence.
The project is being built in public, under the MIT license, one release at a time.
If this direction interests you, explore the repository, test the package, share feedback, or contribute.
RagLeap Voice — Open. Local. Self-hosted. No vendor lock-in.
Top comments (0)