{"categories":[{"label":"Artificial Intelligence","url":"https://skillfed.io/packages/category/scientific-engineering-artificial-intelligence/8"}],"enrichment":{"capability":"Decodes CTC (Connectionist Temporal Classification) logits from speech recognition models using beam search with optional n-gram language model support, handling both standard and byte-pair-encoded vocabularies.","skillfed_tags":["speech-recognition","beam-search","language-models"],"use_cases":["Improve accuracy of Conformer-CTC or Wav2Vec2 models by integrating a domain-specific language model during inference without retraining.","Boost recognition of important domain terms (product names, technical jargon) by adding hotwords during decoding.","Extract word-level timing and confidence scores from CTC models natively through the decoding process.","Process speech recognition logits in batch with multiprocessing for production inference pipelines.","Decode byte-pair-encoded vocabularies automatically without manual token merging logic."],"what_it_does":"pyctcdecode is a Python-based CTC beam search decoder for converting raw logits from speech recognition models into text. It implements shallow fusion with n-gram language models (via kenlm), allowing you to improve transcription accuracy by integrating domain-specific language constraints during decoding rather than retraining the model. The package handles both standard character-based and byte-pair-encoded (BPE) vocabularies automatically, and supports features like hotword boosting (promoting specific phrases during inference) and real-time stateful decoding for streaming applications.\n\nThe decoder is designed as a middle ground between C++ performance and Python flexibility. It includes batch processing via multiprocessing, returns detailed beam search results with frame-level timing information, and integrates with pretrained models from Nvidia NeMo and Hugging Face Wav2Vec2. Runtime is comparable to C++ implementations at beam widths of 10 or greater. However, the package is no longer maintained\u2014the last release was in January 2023 and the repository shows no commits since July 2023.","worth_installing":"Yes, if you have an existing CTC-based speech model and need to integrate a language model or extract frame-level timing. The low install friction and permissive license make it straightforward to add. However, be aware the package is abandoned\u2014no updates since early 2023\u2014so use it only in stable production contexts where you can maintain it yourself if needed. For new projects, verify whether active alternatives exist."},"id":"pyctcdecode","links":{"html":"https://skillfed.io/packages/pyctcdecode","md":"https://skillfed.io/packages/pyctcdecode.md","pypi":"https://pypi.org/project/pyctcdecode/"},"maintenance":{"status":"abandoned"},"meta":{"latest_release":"2023-01-20","license_spdx":null,"license_treatment":"permissive","name":"pyctcdecode","python_support":"supports_current","summary":"CTC beam search decoder for speech recognition."},"popularity":{"monthly_downloads":143734,"position":11168,"tier":"top_15000"},"security":{"n_vulnerabilities":0},"version":"0.5.0"}