feat(mistralai): add streaming STT support for Voxtral Realtime

[zkewal]

Summary

Adds realtime streaming transcription to livekit-plugins-mistralai using Mistral's voxtral-mini-transcribe-realtime-2602 model via the first-party SDK's RealtimeTranscription API. The existing batch transcription path is preserved unchanged.

Why This Matters

The realtime model streams partial transcription deltas mid-utterance instead of waiting for the full audio, reducing perceived latency. It also addresses the batch model's tendency to hallucinate on silence significantly from our private evals. See #4754 for context.

Changes

stt.py (+470/-12)

• Add SpeechStream(stt.RecognizeStream) — fully async streaming via RealtimeTranscription.transcribe_stream() with 50ms PCM audio chunks
• Debounced finalization after VAD FlushSentinel (finalize_delay_ms, default 100ms), plus idle-based finalization (650ms) for the STTNode path where FlushSentinel is not used
• Map Mistral events to LiveKit lifecycle: TextDelta → INTERIM_TRANSCRIPT, SegmentDelta/Done → FINAL_TRANSCRIPT, with START/END_OF_SPEECH
• Unknown event recovery and duplicate final detection
• Model-based routing: batch models use recognize(), realtime models use stream(). Calling recognize() with a realtime model raises a clear error
• _build_capabilities() with graceful fallback for offline_recognize across livekit-agents versions

models.py

• Add voxtral-mini-transcribe-realtime-2602 to STTModels

pyproject.toml

• mistralai>=1.9.11 → mistralai[realtime]>=1.12.0 (adds WebSocket support)
• livekit-agents>=1.4.1 → >=1.3.5 (no 1.4-specific APIs used; avoids OpenTelemetry version conflict with mistralai==1.12.0)

__init__.py

• Export SpeechStream

Usage

from livekit.plugins.mistralai import STT

# Streaming (new)
stt = STT(model="voxtral-mini-transcribe-realtime-2602")
stream = stt.stream()

# Batch (unchanged)
stt = STT(model="voxtral-mini-latest")
event = await stt.recognize(buffer)

Testing & Validation

• Verified batch backward compatibility (existing models unaffected)
• Validated VAD flush finalization and idle-timeout finalization paths
• Confirmed no dependency conflicts with livekit-agents>=1.3.5

Closes #4754

[zkewal]

@chenghao-mou there is a dependency conflict with the mistralai first-party SDK, which includes web-socket support for the streaming model and livekit/agents, over the opentelemetry-api package. The livekit/agents repo requires version >=1.39.0, which was bumped with the 1.3.6 release, while the mistralai Python SDK depends on version 1.38.0.

How do you suggest I handle this? In my local testing, I downgraded livekit/agents to 1.3.5 and was able to test the current PR.

More details here: https://github.com/livekit/agents/actions/runs/21839471789/job/63019576507?pr=4773#step:5:13

submitted an issue to Mistral AI to raise the upper limit: mistralai/client-python#341