Support SPEECH TO TEXT capability

[didux123]

Hey, it would be great to have son STT capabilities !

Thanks !

[Kamilbenkirane]

Design Spec for STT — celeste.audio.transcribe()

Operation.TRANSCRIBE already exists in the enum (src/celeste/core.py) but is unused — no models, no providers, no client methods wire it up. Here's a proposal.

Provider Landscape (verified)

Provider	API	Protocol	Models	Streaming	Notes
OpenAI	/v1/audio/transcriptions	REST	whisper-1, gpt-4o-transcribe, gpt-4o-mini-transcribe	Partial (stream=True)	Speaker diarization via gpt-4o-transcribe-diarize
Google	Cloud Speech-to-Text	gRPC (streaming) / REST (batch)	chirp, chirp_2, latest_long/short	Yes (gRPC only)	Enterprise-grade, 10MB streaming limit
Groq	/v1/audio/transcriptions	REST (OpenAI-compatible)	whisper-large-v3, whisper-large-v3-turbo	No	216x real-time speed
ElevenLabs	/v1/speech-to-text	REST	Scribe v2, Scribe v2 Realtime	Partial (WebSocket option)	99 languages, 150ms realtime latency

What already exists

• Operation.TRANSCRIBE in src/celeste/core.py
• Audio modality with providers (OpenAI, Google, ElevenLabs, Gradium)
• Multi-operation pattern proven in Images (generate + edit) and Text (generate + analyze)
• AudioInput, AudioOutput, AudioChunk types in src/celeste/modalities/audio/io.py

Architecture notes

Output type mismatch: Current AudioOutput wraps AudioArtifact (bytes). STT output is text, not audio. This needs either:

• A new output type (e.g., TranscriptionOutput wrapping str)
• Or routing through the Text modality via DOMAIN_OPERATION_TO_MODALITY

Missing wiring:

• DOMAIN_OPERATION_TO_MODALITY needs (Domain.AUDIO, Operation.TRANSCRIBE) entry
• AudioStreamNamespace / AudioSyncNamespace need transcribe() methods
• AudioInput needs optional audio field (currently only has text: str for TTS)

Proposed API

import celeste

# Basic transcription
output = await celeste.audio.transcribe(
    audio="path/to/audio.mp3",  # or AudioArtifact, bytes, URL
    model="whisper-1",
    language="en",
)
print(output.content)  # str

# Streaming transcription (where supported)
async for chunk in celeste.audio.stream.transcribe(
    audio="path/to/audio.mp3",
    model="gpt-4o-transcribe",
):
    print(chunk.content, end="")

# With timestamps
output = await celeste.audio.transcribe(
    audio="path/to/audio.mp3",
    model="whisper-1",
    timestamps=True,
    output_format="verbose_json",
)

Implementation phases

1. IO types + wiring (TranscriptionOutput, AudioInput.audio field, domain mapping)
2. OpenAI Whisper provider (most straightforward)
3. Groq provider (OpenAI-compatible, thin wrapper)
4. ElevenLabs Scribe v2 provider
5. Google Cloud Speech-to-Text (gRPC — more complex)

Not in scope

• Real-time streaming STT (live microphone → text) — covered by feat: Real-time bidirectional voice — Operation.CONVERSE #201
• Speaker diarization (defer to future)
• Translation (Whisper /v1/audio/translations — separate feature)

[Kamilbenkirane]

@didux123 we're on it!