feat: audio transcribing

README.md

@@ -1,6 +1,6 @@
 # Telegram OpenAI-Compatible Bot
 
-Minimal aiogram bot that forwards whitelisted Telegram users' text messages to an OpenAI-compatible chat completions API.
+Minimal aiogram bot that forwards whitelisted Telegram users' text and voice messages to an OpenAI-compatible API.
 
 ## Features
 
@@ -11,6 +11,7 @@ Minimal aiogram bot that forwards whitelisted Telegram users' text messages to a
 - Persistent per-user context in SQLite
 - `/new` command to reset the current user's context
 - Per-user message batching before sending to the model
+- Voice message transcription through a separately configured transcription model
 - Model-controlled multi-message replies with `<break>` tags
 - Configurable model and system prompt
 - API timeouts, retries, response limits, and safe user-facing fallbacks
@@ -37,6 +38,7 @@ BOT_TOKEN=123456:telegram-token
 API_BASE_URL=https://api.openai.com/v1
 API_TOKEN=sk-your-api-token
 MODEL=gpt-4o-mini
+TRANSCRIPTION_MODEL=whisper-1
 WHITELIST_USER_IDS=123456789
 DATABASE_PATH=data/bot.sqlite3
 MAX_CONTEXT_MESSAGES=20
@@ -45,6 +47,8 @@ MESSAGE_BATCH_DELAY_SECONDS=3
 
 `API_BASE_URL` must be the API root. Do not include `/chat/completions`; the OpenAI SDK appends that path automatically.
 
+`TRANSCRIPTION_MODEL` is used for Telegram voice message transcription before the transcript is sent to `MODEL`.
+
 For OpenRouter, use:
 
 ```dotenv
@@ -75,6 +79,12 @@ The bot stores successful user/assistant turns in SQLite at `DATABASE_PATH`. Eac
 
 Send `/new` to delete your stored messages and start a fresh conversation.
 
+## Voice Messages
+
+Telegram voice notes are downloaded, transcribed with `TRANSCRIPTION_MODEL`, then the transcript is passed through the same batching and chat-completions flow as text messages.
+
+Telegram does not expose a bot API to mark a voice note as listened in the user interface, so that step is not supported.
+
 ## Message Batching
 
 `MESSAGE_BATCH_DELAY_SECONDS` delays the API request after each incoming Telegram message. If the same user sends more messages during that delay, the timer restarts and the messages are sent to the model together as separate entries: