DeepSpeech
End-to-end speech recognition engine by Mozilla using TensorFlow.
About
Mozilla's DeepSpeech is an open-source speech-to-text engine based on the architecture described in Baidu's Deep Speech research paper and implemented in TensorFlow. It converts audio to text entirely on-device with no cloud dependency, supports real-time streaming transcription, and runs on hardware ranging from a Raspberry Pi 4 to GPU servers, which made it a popular choice for privacy-preserving and embedded voice applications. Pretrained English models were distributed alongside tooling for training custom models on new languages and acoustic conditions. Development has ended: Mozilla archived the repository in June 2025, making it read-only, so no further updates or fixes should be expected, though the code remains usable under the MPL 2.0 license. It is still referenced by developers studying end-to-end speech recognition or maintaining legacy offline transcription systems.
Should you use DeepSpeech?
Pick it when
Pick DeepSpeech only when you maintain an existing offline system already built on it, or you are studying how end-to-end recognition from the Deep Speech paper was implemented in TensorFlow.
Look elsewhere when
Skip it for any new project: Mozilla archived the repository in June 2025, so no fixes or updates will come. Vosk or Sherpa-ONNX cover the same offline, Raspberry Pi class use case without an archived codebase.
Alternatives to DeepSpeech
- Vosk
Offline streaming on a Raspberry Pi or phone like DeepSpeech, with 20+ languages and bindings for many programming languages instead of an archived, English-first codebase.
- Sherpa-ONNX STT
On-device ASR from k2-fsa for Android, iOS, embedded boards, and WebAssembly, with streaming and non-streaming models, a current replacement for embedded DeepSpeech apps.
- Whisper.cpp
Whisper's 99 languages and translation in a dependency-free C/C++ library, but Whisper works in 30 second windows rather than DeepSpeech's native streaming.
- Moonshine
Tiny 26M to 245M parameter streaming models under MIT in eight ASR languages, where DeepSpeech shipped English pretrained models and is now archived.
Reviews (0)
Leave a Review
No reviews yet. Be the first to review!
Details
- Price
- Free
- Platform
- Local/Desktop
- Difficulty
- Intermediate (3/5)
- License
- MPL-2.0
- Added
- Apr 3, 2026