How to transcribe CRM voice messages on a dual-core CPU without a GPU
Learn how to build a local voice-to-text transcription system for a CRM using faster-whisper, ffmpeg, and cron without cloud APIs or a GPU.

Clients love sending voice messages, but managers in noisy environments find them inconvenient to listen to, and dialogue lists remain uninformative until opened. To solve this problem without relying on external cloud APIs, developers implemented a local voice transcription system running on a small Russian server with just two CPU cores and no GPU. A 35-second voice message is converted into text in approximately nine seconds using optimized open-source tools.
- Developers integrated a local speech-to-text pipeline for messenger voice messages within their sales CRM.
- The system relies on faster-whisper (small model with int8 quantization) running entirely on a CPU.
- Audio decoding is offloaded to ffmpeg to prevent metadata reading errors across various messenger formats.
- Processing is automated via cron, limiting concurrent tasks and preventing task queue buildup using flock.
Choosing the architecture and speech recognition model External cloud transcription services were quickly rejected due to privacy concerns, as client voice messages contain sensitive personal data that should not leave local infrastructure. The developers chose faster-whisper, an optimized implementation of OpenAI's Whisper built on CTranslate2, which delivers significantly faster inference speeds on standard CPUs compared to the original implementation. The processing was isolated on a separate cloud machine featuring only two CPU cores and no hardware acceleration. The system utilizes the small model variant with int8 weight quantization. While conversational Russian speech might occasionally result in minor word errors, the overall context and meaning are accurately preserved, which is entirely sufficient for sales managers. To maximize performance on limited hardware, several parameters were tuned: compute_type='int8' for fast CPU execution, beam_size=1 for greedy decoding on short phrases, and vad_filter=True to automatically strip silence and pauses common in voice notes.
Technical decoding nuances and CRM integration By default, faster-whisper handles file opening through PyAV, but this approach failed when processing raw messenger audio files arriving in various formats like ogg, m4a, and mp3, often triggering metadata reading errors. The developers resolved this by delegating audio decoding entirely to ffmpeg, converting incoming files into a standardized raw float32 array at 16 kHz mono. This decoupled the pipeline from the original file formats, ensuring stability across diverse client devices. Following Unix philosophy, the transcription script operates via standard input and output streams (stdin to stdout), meaning audio files are never written to the server disk, enhancing security. The main CRM server invokes the script via subprocess calls. Automation is handled by cron, paired with flock to prevent overlapping executions and a strict limit on processed messages per run to protect the recognition server from traffic spikes.
Business impact and practical takeaways This local automation approach eliminated the need for external paid cloud APIs while keeping all audio and text data securely on domestic servers. A voice message is automatically transcribed and displayed in the CRM interface roughly one minute after receipt, allowing managers to instantly grasp the customer's intent directly from the chat list without clicking play. The solution proves that practical AI features can be deployed cost-effectively on modest hardware without expensive GPUs. For technical teams building similar internal tools, starting with a small int8 CPU configuration, leveraging ffmpeg for robust audio decoding, and avoiding unnecessary disk storage provides a reliable blueprint. Protecting user data while enhancing operational efficiency makes this local AI integration a valuable asset for modern CRM platforms operating under strict data privacy constraints.
Source: Все публикации подряд на Хабре



