ElevenLabs TTS generates no audio output in phone/voice calls (text/LLM works, audio never plays)

[muhammadrafaykeshavji]

Description

When using ElevenLabs as the Voice/TTS provider (BYOK), the LLM generates a correct text response and the Live Transcript displays it, but no audio is ever played to the caller. This happens both in self-hosted Asterisk ARI telephony calls and in the hosted app.dograh.com Test Audio feature.

Setup

• Voice tab: Provider=ElevenLabs, Model=eleven_flash_v2_5, valid Voice ID (from My Voices collection), valid API key
• Transcriber: Azure Speech Services, language=ur-IN
• LLM: Groq, openai/gpt-oss-120b

Behavior

• Call connects successfully
• LLM responds with 200 OK and correct text (visible in Live Transcript)
• No audio is heard on the call
• After ~30-50s, connection times out and call drops

Logs

PipelineWorker heartbeat warnings repeat every 10s until timeout:
"heartbeat frame not received for more than 10.0 seconds"
Followed by:
"base_output.py:930 | Failed to write audio frame (consecutive failures: 1/10)"

Possible cause: ElevenLabs WebSocket has a 20s inactivity timeout requiring active keep-alive (per their docs). If Dograh isn't sending keep-alive pings during the gap between text generation and audio streaming, the connection may be silently dying before audio is sent.

Environment

• Self-hosted version: v1.41.0 (Docker Compose)
• Also reproduced on app.dograh.com (v1.42.0)

[amaanJvd]

Took a look at this. Two things worth noting: ElevenLabs 8 kHz output is actually handled (output_format_from_sample_rate returns pcm_8000), and the WS TTS service already runs a keep-alive task — so neither the format nor the missing-keepalive theory seems to be it.

The signature that stands out is base_output.py "Failed to write audio frame" plus the pipeline heartbeat never advancing — that reads like the output transport is getting no audio frames at all, i.e. the WS connects and the text goes out, but audio never streams back on the multi-stream context, rather than a format/transport problem. I'd focus on the ElevenLabs multi-stream-input context lifecycle (context-init/flush) and the receive path in pipecat.

Happy to dig into this in the pipecat submodule and help land a fix — is there a repro (a shared ElevenLabs voice/key on a test workflow) I could use to verify? And is the current pinned pipecat commit the one to target?

[arnofrxdd]

I think its because you might have free account, elevenlabs doesnt let you use community voices in free tier accounts, i just checked it using HTTP socket cause websocket was hiding this error "type":"payment_required","code":"paid_plan_required","message":"Free users cannot use library voices via the API.

You can use elevenlabs free voices they work fine