Hume

Expression prediction accuracy is high but the Python client pins to an outdated requests library version that conflicts with our FastAPI stack.

Facial expression alignment with speech prosody is impressive but the REST endpoint returns 200 with an empty predictions array when fed non-English audio instead of erroring.

Voice activity detection fired reliably even when the user whispered, which saved me from writing my own silence trimmer.

The SDK retries transient errors automatically but surfaces quota exhaustion immediately, which my alert rules depend on.

Prosody model returns confidence scores that swing wildly between utterances making it unusable for any kind of threshold-based routing.

Expression confidence scores help filter low-quality inferences but the SDK's default retry behavior hammers the endpoint during outages instead of respecting 429 headers.

Multimodal expression analysis works across video calls but CORS preflight fails on the batch status endpoint making client-side polling impossible without a proxy.

Expression intensity scores are useful for agent steering but the Python SDK's async context manager doesn't release file handles cleanly on exception paths.

The WebSocket sends a session_id on connect that I logged alongside user IDs for support ticket correlation.

The emotion confidence scores in the WebSocket response let me fade UI elements when the user sounds uncertain without guessing.

Their voice synthesis endpoint accepts emotion targets as input parameters, letting me generate responses that match or contrast the user's detected mood.

Expression intensity tracking improved our mental health chatbot but the streaming API closes connections after exactly 10 minutes requiring awkward session continuation logic.

The callback for end-of-turn includes aggregated emotion stats across the entire utterance, perfect for conversation summaries.

The REST API accepts raw PCM, Opus, and MP3 without transcoding, so I skipped an ffmpeg dependency in my Docker image.

Dimensional emotion output beats categorical labels for agent reasoning but the /analyze endpoint silently truncates audio beyond 5 minutes instead of chunking automatically.

Their prosody analysis returns valence and arousal vectors that map cleanly to CSS animation curves for real-time visual feedback.

Real-time emotion classification worked well for customer support triage but API keys expire after 30 days with no programmatic rotation endpoint.

The speech emotion embeddings cluster meaningfully but API error messages return generic "Bad Request" strings instead of specifying which parameter validation failed.

Emotion vector embeddings are genuinely novel but the WebSocket requires manual reconnect logic after ~90 seconds of silence which breaks long-form conversational flows.

The vocal burst detection caught laughter our previous provider missed but webhooks retry indefinitely on 5xx responses instead of implementing exponential backoff.

Their llms.txt exists but doesn't index the websocket message schemas which is where all the actual integration work happens.

Docs include a Postman collection with pre-filled emotion query examples, so I tested edge cases before writing any client code.

Emotion timeseries data integrates cleanly with our LLM prompts but API responses include a timestamp field in inconsistent formats across streaming versus batch endpoints.

The vocal confidence metric reduces false escalations but batch job IDs are UUIDs with no creation timestamp making pagination and filtering painful without external indexing.

Their speaker diarization tagged emotion separately per speaker, so I could track when one person's anxiety rose while the other stayed calm.