Voice feeling very robotic and less expressive

Telephony
, , , , , ,
1

Hi I am building a telephony outbound voice agent, and I am using sarvam’s STT and TTS models but the voice is feeling very robotic and without expressions.

It’s like I am talking to a machine and also I want to make it multilingual as well, so please guide me on these topics.

2

Unfortunatly, Sarvam doesn’t support SSML, which would offer more expressiveness. If you want to continue with that TTS provider. I’d suggest a few places to look:

  • Confirm your model. bulbul:v3 for new agents. v3 also gives you temperatur (default 0.6)
  • Start at pace=1.0, temperature=0.6, and change one setting at a time.
  • Shorten what reaches TTS. Long LLM paragraphs get chunked awkwardly and lose natural phrasing. Prompt for short conversational sentences, tune min_buffer_size / max_chunk_length.
  • Speaker setting. bulbul:v3 has around 30 voices vs. 7 on v2, and there’s real variation in warmth between them.
  • Check Audio format against your telephony path. Default is 22050 Hz, and the docs say to use speech_sample_rate=8000 only if your SIP provider actually requires narrowband — in which case set output_audio_codec to mulaw or alaw to match. A mismatch that gets resampled downstream will make any voice sound synthetic.
  • Use dict_id (v3 only) if names, brands, or product terms are getting mispronounced on calls.

For multilingual, set target_language_code explicitly rather than relying on defaults, and pair it with the Sarvam STT plugin. The thing to plan for early: the text you send to TTS has to match both the language and the script, so if your LLM replies in romanized Hinglish while you’re set to hi-IN, output degrades. Decide whether you’re normalizing to Devanagari or staying on en-IN for code-mixed speech, and put that in the system prompt.

If it’s still too flat after all that — you can look at Azure or Google Indic voices, or dropping the pipeline for a realtime speech-to-speech model. Worth trying the tuning first though.

Good luck.

3

@Tanish_Rajput Did this response help?

4

Thanks for the detailed suggestions. I’m already using bulbul:v3 with the Priya speaker, pace = 1.0, and temperature = 0.7, but the voice still feels a bit robotic. I’ll definitely look into the audio format, dict_id, and shortening the text sent to TTS as you suggested. Really appreciate your valuable insights—thanks again!