Application for analyzing call transcriptions
Audio Processing
A local application that transcribes phone call recordings of sales managers (RU/UK, particularly mixed language) and automatically evaluates the quality of the conversation based on 8 weighted criteria — from greeting to closing the deal — with verifiable quotes and protection against model "hallucinations."
Key features:
Transcription in two ways: locally (faster-whisper, without sending data anywhere) or via OpenAI API, with a comparison mode for results side by side
Auto-detection and preservation of the original language (Ukrainian/Russian/mixed) without translation
Call evaluation based on 8 criteria with customizable weights (conversation start, politeness, needs detection, active listening, presentation, price, handling objections, closing) — weights and checklists are extracted into a YAML config for a specific business
Each model remark is supported by a quote from the transcript with a timestamp; a separate validator programmatically checks that the quote is indeed in the text — if not, the evidence is marked as unverified and not taken at face value
Support for company profile (prices, products, objection scripts) — the model cannot claim that the manager "named the correct price" if there is no benchmark data
Level of analysis reliability (high/medium/low), which decreases with poor recording quality, lack of role separation between manager/client, short call
Analysis history in SQLite, export to Markdown/HTML/JSON/CSV
Key features:
Transcription in two ways: locally (faster-whisper, without sending data anywhere) or via OpenAI API, with a comparison mode for results side by side
Auto-detection and preservation of the original language (Ukrainian/Russian/mixed) without translation
Call evaluation based on 8 criteria with customizable weights (conversation start, politeness, needs detection, active listening, presentation, price, handling objections, closing) — weights and checklists are extracted into a YAML config for a specific business
Each model remark is supported by a quote from the transcript with a timestamp; a separate validator programmatically checks that the quote is indeed in the text — if not, the evidence is marked as unverified and not taken at face value
Support for company profile (prices, products, objection scripts) — the model cannot claim that the manager "named the correct price" if there is no benchmark data
Level of analysis reliability (high/medium/low), which decreases with poor recording quality, lack of role separation between manager/client, short call
Analysis history in SQLite, export to Markdown/HTML/JSON/CSV