Podcast generation failed twice at the audio stage, for two unrelated reasons
that look similar from the UI but need opposite fixes.
1) TTS/STT ran as nohup background processes and silently did not survive a
reboot. Podcasts then failed with "Failed to generate speech: All connection
attempts failed" (httpx.ConnectError) even though outline and transcript had
generated fine.
Both now run as systemd user units. The unit files are versioned in
services/systemd/ (using %h, not a hardcoded home) and installed by
scripts/start_services.sh, which also enables linger so they start on boot
without a login session. They pin GPU 2 by UUID, not by index: CUDA orders
devices "fastest first", so index 2 can resolve to the T600.
2) GPU 2 is shared by three processes (TTS, STT and the separate chatterbox-tts
MCP service on :9999), leaving ~12 GB of headroom. podcast_creator sends
TTS_BATCH_SIZE (default 5) clips concurrently, and since /audio/speech is a
sync FastAPI handler, they generated genuinely in parallel on one shared
model. Activation memory multiplied, the TTS process hit 16.7 GB and threw
torch.OutOfMemoryError, surfacing as "HTTP 500" from the endpoint.
tts_server.py now serializes generation behind a lock (GPU-bound work, so
parallelism buys no throughput — it only multiplies peak VRAM) and frees the
cache afterwards. TTS_BATCH_SIZE=1 keeps the client from queuing requests in
that lock and running into esperanto's 300s TTS timeout; ESPERANTO_TTS_TIMEOUT
is raised to 600s as headroom.
Verified: 5 concurrent /audio/speech requests all return 200 with GPU 2 peaking
at ~11.5 GB (was 16.7 GB for the TTS process alone), and the previously failed
episode now completes end to end — 38/38 batches, 10:56 min of audio, zero OOM.
Docs record both failure signatures side by side, since ConnectError (server
dead) and HTTP 500 (server alive, out of VRAM) have very different remedies.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Two root causes, both fixed in the bind-mounted podcast prompt templates
(prompts/podcast/*.jinja):
1. Language: the episode profile's `language` field IS passed to both
podcast_creator templates as {{ language }}, but the stock templates never
reference it — so podcasts came out English (driven only by the English
stock briefings/speakers), then read aloud by the German-locked Chatterbox
TTS = "English with a German accent". Added a CRITICAL LANGUAGE REQUIREMENT
block keyed on {{ language }} to both templates; now `language: "de"`
actually forces German. Verified end-to-end: a full run produced a 44-line
all-German transcript + audio.
2. Invalid json output failures: qwen3.5:27b is a thinking model; on long
segments the <think> block ate the response-token budget and truncated the
JSON. Prepended /no_think to both templates (~3x faster, valid JSON).
Templates are bind-mounted read-only via docker-compose (directory mount, so
edits survive a container restart without inode-staleness). Bundled briefings
and speaker backstories were also translated to German to reduce drift.
Known limitation documented: feeding an entire book (~70k tokens) as podcast
content makes each of the 6 LLM calls take ~3.5 min and is unreliable; use a
shorter source or summary. Confirmed working with concise content.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
