Price: 100000
Number of applications: 0
12.11.26 (inclusive)
Cash payment
Idea
ICT tasks
Robotics
Neurotechnology and artificial Intelligence
Purpose and description of task (project)
The goal of the project is to create an automatic transcription service that converts audio recordings of calls into text. The service operates as a standalone tool: a file is transmitted via REST API or uploaded manually, and in response, a ready-made text transcription is returned. The service is only responsible for speech recognition. It does not retrieve recordings from storage, and the obtained text is not saved in knowledge bases or transferred to other systems. The service should accept audio in common formats (WAV, MP3, OGG/Opus, GSM) and recognize speech in Kazakh and Russian, including mixed speech. The result includes text with punctuation, speaker segmentation, and timestamps. It can be obtained in TXT, JSON, SRT, and DOCX formats. The solution is deployed entirely within the customer's infrastructure (on-premise). It should quickly process recordings of varying lengths and quality, including telephone quality, and easily connect to the company's existing systems without major modifications.
Software/ IS
Call center recordings accumulate but are almost never used: to understand a conversation, it has to be listened to in full, and manually transcribing one hour of recording takes an employee several hours. Typical recognition systems poorly understand Kazakh and mixed speech, telephone sound, and medical terms. Cloud services cannot be used because the recordings contain personal and medical data of patients. The service should accept audio files, recognize them without human involvement, and return a text transcript. Files are uploaded via REST API or manually through the web interface in WAV, MP3, OGG/Opus, and GSM formats, mono and stereo. Tasks are queued with retries, and their status can be checked via API. The service recognizes Kazakh and Russian speech, including mixed, and identifies the language for each segment. It separates the operator's and client's remarks, adds punctuation, normalizes numbers and dates, and takes into account a dictionary of medical terms. The result is returned in JSON with segments, speakers, timestamps, and confidence scores, and can be exported in TXT, SRT, or DOCX. The solution operates entirely within the customer's infrastructure: neither the audio nor the text is transmitted to third parties. Target metrics for telephone recordings are as follows: recognition accuracy (WER) no worse than 15% for Russian and 25% for Kazakh; accuracy of speaker separation (DER) no worse than 15%. One hour of recording should be processed in 6 minutes on a single GPU, and throughput should be at least 2,000 hours of audio per month. The result is ready no later than an hour after upload, with 99% availability during working hours. The solution complies with the Law of the Republic of Kazakhstan "On Personal Data and Their Protection," and the data is stored within the territory of Kazakhstan. Access is only available to authorized users, traffic and data are encrypted, personal data in the text can be masked, and the storage period for recordings is adjustable.
After the service is implemented, the time required to transcribe calls will be reduced from hours to minutes: one hour of recording will be processed in approximately 6 minutes. The costs of manual transcription and listening will significantly decrease, and employees will be able to focus on their core tasks. All calls will be transcribed, not just selective ones. All transcriptions will be output in a unified format with clear quality, speaker labeling, and timestamps. It will become easy to work with the call texts: reading them, finding necessary fragments, and checking the content. The processing will remain entirely within the customer's environment, ensuring compliance with personal and medical data protection requirements. The service will allow processing large volumes of audio without hiring additional staff, and its API will serve as the foundation for further automation: search, analytics, and quality control.
Шахворостова Маргарита Николаевна
Whatsapp
