Mishear a single letter in a drug name, and even the most natural voice agent is useless

A typical voice agent pipeline has three stages: STT converts the caller's audio into text, the LLM processes the text and generates a reply, and TTS turns the reply back into audio. Akshay Pachaar lays out the problem clearly: information can be lost at two boundaries. STT has to preserve what the caller said, and TTS has to preserve what the LLM generated. If the caller says RX-B17, STT may hear RX-D17; if the LLM correctly generates RX-B17, TTS may still pronounce it as RX-D17. Drug names, dosages, appointment times, tracking numbers, and alphanumeric codes all get stuck at these two boundaries.

Deepgram built Flux TTS for the voice layer. The evaluation Akshay cites includes drug names, tracking numbers, amounts, technical strings, and alphanumeric strings, and Flux recorded a 3.4% word error rate. He also ran a complete voice loop: a veterinary clinic scenario, renewing a prescription for a dog named Biscuit. The agent had to understand and repeat back the medication, dosage, prescription number, and a pickup time that changed during the call. The test result was that the details were preserved, the voice sounded natural, and the interaction felt like an acceptable veterinary clinic phone call. Akshay notes that this content was made in collaboration with the Deepgram team, with links and credit promotion at the end.

The video is a full recording of Akshay's test:

Build link: Flux TTS. Deepgram's promotion matches $1 of credit for every $1 spent on Flux TTS, up to $500.

3.4% looks low, but it isn't good enough for entity strings. In a reply, Aakash Verma points out that WER averages over each word, so getting one character wrong in RX-B17 barely moves the metric; this scenario should separately measure exact match for codes, dosages, and times. seridarivus adds another layer: STT and TTS errors on codes sound perfectly fluent, and only read-back confirmation can catch what WER hides. One wrong character and the entire prescription number is useless.

Phone lines themselves aren't clean either. Thomas says B and D sound alike on bad lines, a problem that plagued pharmacy phone calls long before LLMs existed. Vibefold offers a cheap fix: before acting, the agent should read back anything that looks like an ID or number. Kalu goes into more detail: read it character by character, r x b one seven, and ask the other person to confirm. Jean Hyacinthe's approach is to have the caller type in the highest-risk callback number on the keypad, then read it back for confirmation; when they used Haitian Creole, dictated numbers kept failing at the STT stage. Champ warns that tests should have the caller change the dosage mid-call—for example, 5mg rather than 50—and then check what the agent wrote down, not just what it says in the audio.

Latency is another issue. Agentik says the three hops add up faster than you'd think, and in their timing, waiting for TTS's first audio chunk took longer than the LLM call. One commenter flatly concludes that the three-stage setup will always be slower and is not optimistic about this path. Sunil's take is closer to how users feel: callers forgive a wrong word faster than they forgive a two-second pause. Someone in the video comments asks why the speech sounds like it's at 2x speed, and both TaraT and Jacki Maniel question how you can judge real performance after speeding it up. Demos can be edited, but speech rate and latency should be clearly labeled.

Pritesh Kumar's summary is a good place to close: three stages, two places where information can die; STT can drop or alter codes, TTS can mispronounce content the LLM already generated; swapping in a better model on one side does not automatically protect the other side. As voice AI enters business, don't first compare who sounds more human; first compare who makes one fewer character error. Get the dosage, time, or number wrong once, and even the most natural voice won't earn trust. That's also what Jingge means in the reply.