Audio transcription (speech-to-text)

LLM.transcribe() converts audio into text. Not supported by every provider — it depends on whether the provider's API accepts audio:

Provider	Supported?	How	Typical models
OpenAI	✅	dedicated endpoint	`gpt-4o-transcribe`, `gpt-4o-mini-transcribe`, `whisper-1`
Groq	✅	dedicated endpoint	`whisper-large-v3`, `whisper-large-v3-turbo`
Gemini	✅	multimodal (`generateContent`)	`gemini-2.5-flash`, `gemini-2.5-pro`
Anthropic	❌	—	the Claude API does not accept audio

Trying to transcribe on Anthropic raises UnsupportedError. Claude's "voice" is a product feature (app/Claude Code), not part of the model API.

Usage

from jangada_ai import LLM, Audio

# OpenAI
llm = LLM("openai", "gpt-4o-transcribe")
print(llm.transcribe("interview.mp3").text)

# Groq (faster/cheaper)
llm = LLM("groq", "whisper-large-v3-turbo")
print(llm.transcribe("interview.mp3").text)

# Gemini (multimodal)
llm = LLM("gemini", "gemini-2.5-flash")
print(llm.transcribe("interview.mp3").text)

Accepted inputs: a path, an AudioPart, or bytes via Audio.from_bytes:

audio = Audio.from_bytes(blob, "audio/wav", name="speech.wav")
llm.transcribe(audio)

Async: await llm.atranscribe(audio).

Options per provider

Extra kwargs go straight to the provider when it accepts them:

# OpenAI/Groq accept language, prompt, response_format, temperature, ...
llm.transcribe("audio.mp3", language="pt", response_format="text")

# Gemini accepts prompt= as an instruction (e.g. include timestamps)
llm.transcribe("audio.mp3", prompt="Transcribe with timestamp markers.")

Fallback between audio providers

Like any jangada call, transcribe() honors retry and fallback:

from jangada_ai import LLM

primary = LLM("groq", "whisper-large-v3-turbo")
backup  = LLM("openai", "gpt-4o-transcribe")
stt = primary.with_fallback(backup)

stt.transcribe("audio.mp3")   # tries Groq; if it fails (5xx/timeout), goes to OpenAI

UnsupportedError does not enter the default failover — it's a configuration error (provider without audio), not a transient failure. See Errors and Retry and fallback.

Formats and limits

OpenAI/Groq: flac, mp3, mp4, mpeg, mpga, m4a, ogg, wav, webm — up to ~25 MB.
Gemini: inline audio up to ~20 MB for the whole request (use the SDK's Files API for larger files).

Example

examples/transcribe_example.py — runnable script.