Audio & Turn ProblemsCopy Anchor Link

Most "my agent feels robotic or buggy" feedback on live phone calls comes down to Turn / VAD tuning or TTS pronunciation, not the prompt itself. Match what you hear on a test call to one of the playbooks below.

---

Playbook 1: "The bot cuts the caller off mid-sentence"Copy Anchor Link

Likely CausesCopy Anchor Link

• Silence Before Responding (turn_silence_timeout_secs) is too low (0.3–0.6s), causing brief thinking pauses or digit-by-digit speech to trigger a bot reply.
• Detection Strategy (turn_stop_strategy) is set to Transcription-based (transcription) on an open-ended question where callers frequently trail off ("um, actually...").

Fix StepsCopy Anchor Link

1. Raise Silence Before Responding (turn_silence_timeout_secs) to 1.0–1.4s (valid UI range 0.3–2.5s; Balanced preset uses 0.9s).
2. Or switch Detection Strategy to Smart Turn Analyzer (turn_analyzer) and set Incomplete Turn Timeout (smart_turn_stop_secs) to 2.0–2.5s (UI range 0.5–10s, default 2.0s).
3. Keep VAD Silence Hold (vad_stop_secs) short (0.20–0.35s, default 0.25s)—tune conversational pacing via turn_silence_timeout_secs, not vad_stop_secs.

---

Playbook 2: "Road noise / office chatter / TV interrupts the bot while it's speaking"Copy Anchor Link

Likely CausesCopy Anchor Link

• Volume Floor (vad_min_volume, default 0.20) and Neural Confidence (vad_confidence, default 0.50) from the High Sensitivity defaults are too permissive for noisy caller environments.
• Interruption Strategy (turn_start_strategy) is on Default (default), which interrupts immediately on any speech frame.
• On the opening greeting or compliance script, Allow Interruption (allow_interrupt) was enabled.

Fix StepsCopy Anchor Link

1. Switch the Voice Sensitivity Preset to Noisy Room (vad_min_volume = 0.55, vad_confidence = 0.75, turn_silence_timeout_secs = 1.2s, user_turn_stop_timeout = 5.0s).
2. Set Interruption Strategy to Minimum words (min_words) with 2 or 3 words (default 3), or test Provisional VAD (provisional_vad, 1.0–1.5s) so coughs and short background bursts don't kill bot playback.
3. On Start Call (default false) and mandatory compliance Agent nodes, set Allow Interruption to OFF (false).
4. Keep Ambient Background Audio volume modest (default 0.3, or lower/disable if callers report clutter).
5. (Remember: Realtime openai, azure, and grok S2S models use provider server VAD and ignore local Silero VAD settings.)

---

Playbook 3: "The bot doesn't hear quick 'yes', 'ok', or 'haan' replies"Copy Anchor Link

Likely CausesCopy Anchor Link

• Noisy Room preset (vad_min_volume = 0.55, vad_confidence = 0.75) is filtering out soft whispers or compressed 8 kHz landline mics.
• Boost Short Affirmations (boost_affirmations) was turned off, or STT is dropping 1-syllable answers.

Fix StepsCopy Anchor Link

1. Switch to High Sensitivity preset (Volume Floor 0.20, Neural Confidence 0.50, Silence Wait 0.7s, Watchdog 3.5s).
2. Ensure Boost Short Affirmations is ON (true) in Turn Detection settings.
3. Keep vad_stop_secs near 0.25s and keep min_words off (Default) on rapid yes/no confirmation flows.

---

Playbook 4: "There is a 3–4 second awkward silence after the caller finishes speaking"Copy Anchor Link

Likely CausesCopy Anchor Link

• STT didn't finalize a short 1-word turn and the Safety Watchdog Timeout (user_turn_stop_timeout) had to expire (3.5s local default, 4.0s external STT, up to 5.0s on Noisy Room preset).
• Deepgram Flux (or Cartesia ink-2) external turn ownership is active with high provider EOT timeouts, bypassing local turn_silence_timeout_secs.
• A Mid-Call HTTP Tool (http_api) is blocking before generating speech without a customMessage filler.

Fix StepsCopy Anchor Link

1. Check Silence Before Responding (turn_silence_timeout_secs) (0.7–0.9s for snappy turns) and lower Safety Watchdog Timeout (user_turn_stop_timeout) toward 2.5–3.0s, while ensuring Boost Short Affirmations is ON.
2. Verify Production Provider Tiering: prefer Tier-1 providers (OpenAI gpt-4.1-mini / gpt-4.1-nano / gpt-4o-mini, Deepgram Nova-3 or Sarvam saaras:v4 for STT, Sarvam bulbul:v4-flash or SmallestAI lightning_v3.1 for fast telephony TTS).
3. If an HTTP tool takes > 1s, add a short spoken customMessage on the tool and move slow multi-step CRM logic behind https://n8n.talkar.in.

---

Playbook 5: "The bot mispronounces brand names, Indian numbers, or dates"Copy Anchor Link

Likely CausesCopy Anchor Link

• Numbers/codes (1042, ORD-902) are written unformatted in prompts or returned raw by tools.
• Recognition vocabulary hints weren't added to the workflow Dictionary (or exceeded provider caps).

Fix StepsCopy Anchor Link

1. Use Workflow Dictionary (Configurations -> Dictionary): Add comma-separated brand names, SKUs, and Indian names (50–100 limits vary: Deepgram Nova-3 supports up to 100 keyterms, Sarvam caps at 50 terms max 64 chars on saaras:v4, and SmallestAI uses its own weighted STT keywords field up to 100 rather than the workflow dictionary).
2. Format Tool Responses: Return human-speakable strings from n8n.talkar.in (e.g., "two thousand four hundred rupees" or "O R D 9 0 2" instead of "INR2400" or "#ORD902").
3. Add Prompt Formatting Rules: Instruct the Agent node to spell out order IDs character-by-character and output Latin script only if Hinglish callers are being answered by an English voice.
4. Test Sarvam (saaras:v4 / bulbul:v4-flash) vs. SmallestAI (pulse / lightning_v3.1) on a real PSTN Direct Test Call (8 kHz).