Hacktoberfest 2026: the issues maintainers tagged for October, open and beginner-friendly. Browse Hacktoberfest issues

MagpieTTS long-form: a 4–5 character sentence fails the whole request ("longform history context cache is too short")

Open
  • Is something wrong?
#61 0 comments 0 reactions 0 assignees View on GitHub

Maintainers usually reply within 3 days

Nobody has claimed this yet.

Assessment

Difficulty
4/5
Estimated time
3-5 days
Newbie friendliness
52/100
Issue type
Bug
Clarity
Mostly clear
Activity status
Active
Tech stack
cpp
Domain
audio-video-rtc, backend

Research direction

Start by locating the MagpieTTS long-form synthesis path and the history-context cache check that reports “need 20 token(s)”; the issue does not name source files or tests. Reproduce with the provided command and compare long-form behavior for very short sentences versus the workaround. Done means short sentences no longer fail the request, while long-form output is not silently truncated.

Written by the indexing model from the issue text.

Description

Version: nemo-speech 0.1.0 and 0.2.0 (nemo-speech-0.{1,2}.0-linux-x86_64-cuda
release tarballs; the same input fails identically on both; the nightly was not tried), model nvidia/magpie_tts_multilingual_357m@452ef560f972
(v2602.f16.gguf), codec nvidia/nemo-nano-codec-22khz-1.89kbps-21.5fps.
RTX 3090, driver 570.153.02, Linux 6.8.

What happens: when the input is long enough for long-form mode to engage,
a single very short sentence anywhere in it makes synthesis fail:

[nemo-speech] synthesize session started
longform history context cache is too short: need 20 token(s), have 18
nemo-speech synthesize: MagpieTTS synthesis failed
[nemo-speech] synthesize session failed (exit code 1)

nemo-speech serve returns HTTP 500 "MagpieTTS synthesis failed" for the same
input. It is deterministic: the same text fails every time, with any voice.

Repro: a filler sentence repeated four times, then one short sentence,
then the filler four times again (~900 characters):

F="The service kept running normally while the operators reviewed the dashboards and the alert history in detail."
F4="$F $F $F $F"
echo "$F4 Okay. $F4" > t.txt
nemo-speech synthesize -i t.txt -o o.wav --force --seed 1
# -> exit 1, "need 20 token(s), have 18"

Which sentences fail: the same template, with only the middle sentence
changed:

middle sentence chars result
Why? 4 fails, "have 12"
Yes. 4 fails, "have 12"
Okay. 5 fails, "have 18"
No way. 7 ok
Why not? 8 ok
12 other sentences of 11–20 chars all ok

The same short sentence passes in a short input, where long-form mode does not
engage. In a real 48-chunk text (each chunk ≤ 800 chars), 47 rendered and one
failed. It contained "Why? It ran out of disk space. Why? Logs were not
rotated. Why? …".

--tts.longform options: on behaves like auto (fails). off avoids
the error but truncates silently: an 800-char chunk that gives 53 s of audio
in long-form came back as 23 s, and 31 s with --steps 3000.

Expected: a short sentence is merged with its neighbour, or padded, so that
the history context reaches the minimum. It should not fail the request.

Workaround on our side: before sending text to Magpie, the client attaches
every sentence shorter than 8 characters to the previous sentence (or to the
next one when it comes first): "down. Why?" becomes "down, why?". With that,
the 48-chunk text renders 48 of 48.

Dominant language
C++
Stars
150
Forks
32
Avg merge
7d 17h
Merged PRs (30d)
8

Getting set up

First steps

  1. Read the whole issue, then the project's contributing guide.
  2. Comment on the issue to say you are picking it up — it saves two people doing the same work.
  3. Fork the repository and make your change on a branch.
  4. Open a pull request that references the issue number.

More from NVIDIA/NeMo-Speech.cpp

All issues in NVIDIA/NeMo-Speech.cpp

Similar issues

More C++ issues

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.