Flux TTS: 0.3-2 s pause between two sentences of one utterance, varying run to run #1708
Replies: 4 comments
|
Thanks for asking your question. Please be sure to reply with as much detail as possible so the community can assist you efficiently. |
|
Hey there! It looks like you haven't connected your GitHub account to your Deepgram account. You can do this at https://community.deepgram.com - being verified through this process will allow our team to help you in a much more streamlined fashion. |
|
It looks like we're missing some important information to help debug your issue. Would you mind providing us with the following details in a reply?
|
|
Follow-up with more data, including request IDs as the bot asked. 1. The pause is in the generated audio, not in delivery. We added a per-reply audio log in our pipeline, after the output transport. In one test call, all 8 two-sentence replies had a pause over 700 ms between the sentences, and the longest ran 1.08-1.72 s. Each silence has the noise floor of generated audio: RMS about 8, roughly -72 dBFS, never digital zero. Audio chunks arrived with no gap over 60 ms, and time spent waiting for audio was 0-4 ms per reply. 2. Larger offline sample. We ran 6 different two-sentence texts (an acknowledgement, then a question), 12 runs each, so n = 72 per setting. Each run used a fresh
For Request IDs (
The first and last rows are the same text and settings, a few seconds apart. Is the boundary pause length voice-specific by design? And is there a way to get sienna's pacing closer to brooke's? |
Uh oh!
There was an error while loading. Please reload this page.
We use Flux TTS (
flux-sienna-en,/v2/speakWebSocket, EU endpoint) for a voice interviewer through Pipecat. Most replies are two sentences: an acknowledgement, then a question. Flux often puts a pause of more than a second between the two sentences, and how long it is changes from run to run on the same text. A tester described it as "it feels like 2-3 seconds between the pieces, unnaturally long". A pause like that also invites the user to start talking before the question arrives.Repro
The same two-sentence text was synthesized 8 times streamed (word-sized
Speakmessages, the way an LLM token stream arrives, thenFlush) and 8 times as a singleSpeakmessage. We counted silent runs inside the audio of at least 250 ms, where silent means the peak sample stays under 300 in 10 ms windows. Leading and trailing silence are excluded.Text: "Thanks, that gives me a clear picture of how you handled the handover. What did you do when the first deadline started to slip?"
Silence at the sentence boundary (around 3.2-3.9 s into the audio), on
api.eu.deepgram.com, 2026-10-26:SpeakmessageOn our own call texts we saw up to 2.0 s, and from 0 to 2.0 s across repeats of one text.
Things we tried that did not remove it:
expressivity-2, -1 and 1 (small sample, n=3 each).Script (needs
websockets; setDEEPGRAM_API_KEY):Questions
Thanks!
All reactions