pocketsphinx

Why we included this project

PocketSphinx carries Carnegie Mellon's speech recognition lineage, and despite algorithms that date back decades, it remains a practical choice for offline, on-device transcription without a heavy neural stack. Teams building for constrained hardware like a Raspberry Pi, a voice-controlled robot, or a kiosk that has to work with no network connection will find the C library compiles down to something small and fast, while staying speaker-independent and continuous. The Python bindings make live or file-based prototyping quick, and the command-line tool reads PCM audio from standard input or files, so it drops into existing pipelines. It also does forced alignment at word or phone level, which helps when labeling audio corpora or checking pronunciations. Just keep expectations in check: the models are old, so accuracy trails modern cloud recognizers. Treat it as a lean, dependency-light option when you control the acoustic environment, not a replacement for large-scale ASR.