Audio

Transcribe audio to text and generate speech from text via LiteLLM

Transcribe audio files to text and generate speech from text using any LiteLLM-supported provider (OpenAI Whisper, Deepgram, ElevenLabs, …).

Quick Start

1. Give an Agent a transcription tool

   from praisonaiagents import Agent
   from praisonai.capabilities import transcribe
   
   agent = Agent(
       name="MeetingSummariser",
       instructions="Transcribe the audio, then produce a bullet-point summary.",
       tools=[transcribe],
   )
   agent.start("Transcribe ./meeting.mp3 and summarise the decisions.")

2. Transcribe directly (Whisper default)

   from praisonai.capabilities import transcribe
   
   result = transcribe("./meeting.mp3")
   print(result.text)

3. Generate speech from text

   from praisonai.capabilities import speech
   
   result = speech("Hello, world!", voice="nova")
   result.save("hello.mp3")

4. Use the async variants

   import asyncio
   from praisonai.capabilities import atranscribe, aspeech
   
   async def main():
       transcript = await atranscribe("./meeting.mp3")
       audio = await aspeech(transcript.text, voice="nova")
       audio.save("readback.mp3")
   
   asyncio.run(main())

5. Switch providers without changing code

   result = transcribe("./meeting.mp3", model="deepgram/nova-2", language="en")
   print(result.text)

How It Works

Calls route through LiteLLM to the provider that matches your model string.

Configuration Options

transcribe(...) / atranscribe(...)

speech(...) / aspeech(...)

Result objects

Common Patterns

Transcribe → summarise pipeline

from praisonaiagents import Agent
from praisonai.capabilities import transcribe

transcript = transcribe("./meeting.mp3")

agent = Agent(
    name="Summariser",
    instructions="Summarise the transcript into bullet-point decisions.",
)
agent.start(transcript.text)

Multilingual dubbing

import asyncio
from praisonaiagents import Agent
from praisonai.capabilities import atranscribe, aspeech

async def dub(path):
    transcript = await atranscribe(path, language="es")
    translator = Agent(name="Translator", instructions="Translate Spanish to English.")
    english = translator.start(transcript.text)
    result = await aspeech(english, voice="nova")
    return result.save("dubbed.mp3")

asyncio.run(dub("./clip_es.mp3"))

Word-level timestamps for captions

from praisonai.capabilities import transcribe

result = transcribe(
    "./meeting.mp3",
    response_format="verbose_json",
    timestamp_granularities=["word"],
)
for word in result.words or []:
    print(word)

Best Practices

Related