Massive Speech-to-Text Accuracy Issues over Twilio Calls (German / 8 kHz Narrowband)

Hey everyone,
we’re running into a major issue with STT accuracy on Twilio inbound phone calls (German language).
Our setup:

Twilio PSTN number (8 kHz / G.711) → Voiceflow call flow

Deepgram Nova-2 (DE) as STT model

TTS: ElevenLabs

LLM + Voiceflow Agent for routing (Rezept / Rückruf / Termin)

The problem: the bot often completely misunderstands simple words like “Rezept” or “Rückruf”.
It performs perfectly fine in the browser (wideband mic), but fails badly when going through Twilio.
Example: I clearly say “Ich brauche ein Rezept” → STT returns “Ich komm” or random nonsense.

Already tested:

Deepgram tier: enhanced, smart_format: true, no_delay: false, language: de

AGC active

Google STT V2 as alternative (same issue)

Increased “End of Speech Timeout” to 2000 ms

Clean mic / quiet environment

Still, accuracy stays around 50 %.
I know Twilio’s PSTN audio is 8 kHz narrowband, but is there any known workaround or optimization to improve STT quality?

Questions:

Has anyone found a codec / routing setup (e.g., Opus, PCM16) that actually delivers 16 kHz audio into Voiceflow?

Are there STT models better optimized for telephony audio (8 kHz)?

Would using a Twilio Stream Connector with preprocessing (AGC + noise suppression + echo cancel) help?

Any trick to reduce false triggers or cut-offs (VAD too aggressive)?

Would love to hear if anyone faced this and found a robust solution