Skip to content

Higgs_Audio_TTS voice cloning not working and Higgs TTS generation reached max_tokens before EOC (Arabic, CPU) #471

Closed
Closed
Higgs_Audio_TTS voice cloning not working and Higgs TTS generation reached max_tokens before EOC (Arabic, CPU)#471
Labels
help wantedExtra attention is needed

Description

@user6654126
user6654126
opened on Sep 6, 2026

Environment:

OS:Linux Mint 22.3 - Cinnamon 64-bit
CPU:AMD Ryzen 5 8400F
GPU: RTX 5060 8G VRAM
RAM: 6000MHz 16GB RAM
Model: higgs-audio-v3-tts-4b-q8_0.gguf
version: latest v0.7.2

Symptom:

when i use this model with backend cpu sometime it work , if the TTS script is small meaning less than 2 minutes.
but the problem it doesn't even use reference voice in voice cloning instead it use different and random voices and i noticed it switch into different voice after whatever limit i set in --text-chunk-size 300
so how i make it use the reference voice for voice cloning

Commands arabic test:

./build/linux-cpu-release/bin/audiocpp_cli
--task tts
--family higgs_audio_tts
--session-option "higgs_audio.mem_saver=true"
--model /home/xxxx/audio.cpp/build/linux-cpu-release/bin/models/Higgs-Audio-v3-TTS-4B-GGUF/higgs-audio-v3-tts-4b-q8_0.gguf
--backend cpu
--threads 10
--text "$(cat /home/xxxx/audio.cpp/inputs/gemini-arabic-mini.txt)"
--voice-ref /home/xxxx/audio.cpp/inputs/stoic_10seconds.wav
--reference-text "$(cat /home/xxxx/audio.cpp/inputs/stoic_10seconds.txt)"
--out /home/xxxx/audio.cpp/outputs/higgs_ara_1.wav
--text-chunk-size 300
--metrics
--log

Logs:
higgs_ara_1 log.txt

also if i removed --session-option "higgs_audio.mem_saver=true" \ and --text-chunk-size 300 \ i get Higgs TTS generation reached max_tokens before EOC
and if script is long +5 minutes i get EOC so how to fix that

Commands arabic test:

./build/linux-cpu-release/bin/audiocpp_cli
--task tts
--family higgs_audio_tts
--model /home/xxxx/audio.cpp/build/linux-cpu-release/bin/models/Higgs-Audio-v3-TTS-4B-GGUF/higgs-audio-v3-tts-4b-q8_0.gguf
--backend cpu
--threads 10
--text "$(cat /home/xxxx/audio.cpp/inputs/gemini-arabic-mini.txt)"
--voice-ref /home/xxxx/audio.cpp/inputs/stoic_10seconds.wav
--reference-text "$(cat /home/xxxx/audio.cpp/inputs/stoic_10seconds.txt)"
--out /home/xxxx/audio.cpp/outputs/higgs_ara_1.wav
--metrics
--log

Logs:
higgs_ara_2 log.txt

gemini-arabic-mini.txt
stoic_10seconds.txt
higgs_ara_1.wav
stoic_10seconds.wav

Activity

  1. 0xShug0 commented on Sep 6, 2026

    @0xShug0
    0xShug0
    on Sep 6, 2026
    Owner

    @user6654126 Thanks for the detailed report! I'll reproduce and look into the issue. Quick question: Have you tested english intput?

  2. 0xShug0
    changed the title [-]Higgs_Audio_TTS voice cloning not working and Higgs TTS generation reached max_tokens before EOC[/-] [+]Higgs_Audio_TTS voice cloning not working and Higgs TTS generation reached max_tokens before EOC (CPU)[/+] on Sep 6, 2026

  3. user6654126 commented on Sep 6, 2026

    @user6654126
    user6654126
    on Sep 6, 2026
    Author

    @user6654126 Thanks for the detailed report! I'll reproduce and look into the issue. Quick question: Have you tested english intput?

    I have but i am not interested in using it for English as there are better alternative which is BreezeTTS 2 and i used it successfully using cuda and it's perfect in English so i want the best tts model in Arabic so that's why I am using higgs audio v3 tts for Arabic

  4. 0xShug0
    changed the title [-]Higgs_Audio_TTS voice cloning not working and Higgs TTS generation reached max_tokens before EOC (CPU)[/-] [+]Higgs_Audio_TTS voice cloning not working and Higgs TTS generation reached max_tokens before EOC (Arabic, CPU)[/+] on Sep 6, 2026

  5. 0xShug0 commented on Sep 6, 2026

    @0xShug0
    0xShug0
    on Sep 6, 2026
    Owner

    @user6654126 Thanks for the detailed report! I'll reproduce and look into the issue. Quick question: Have you tested english intput?

    I have but i am not interested in using it for English as there are better alternative which is BreezeTTS 2 and i used it successfully using cuda and it's perfect in English so i want the best tts model in Arabic so that's why I am using higgs audio v3 tts for Arabic

    Sorry I should make it clear: This is just to help me find the root cause. At least a quick test on CUDA on my machine, English and Chinese work correctly without EOC under chunking and produce consisent voices. Arabic EOC on both CPU and CUDA.

    So if English works on your machine on CPU, we can rule out backend/environment quirks. The likely issue is Arabic input processing.

    higgs_english_chunk300_unseeded_cuda.wav

    higgs_chinese_long_chunk300_unseeded_cuda.wav

  6. user6654126 commented on Sep 6, 2026

    @user6654126
    user6654126
    on Sep 6, 2026
    Author

    @user6654126 Thanks for the detailed report! I'll reproduce and look into the issue. Quick question: Have you tested english intput?

    I have but i am not interested in using it for English as there are better alternative which is BreezeTTS 2 and i used it successfully using cuda and it's perfect in English so i want the best tts model in Arabic so that's why I am using higgs audio v3 tts for Arabic

    Sorry I should make it clear: This is just to help me find the root cause. At least a quick test on CUDA on my machine, English and Chinese work correctly without EOC under chunking and produce consisent voices. Arabic EOC on both CPU and CUDA.

    So if English works on your machine on CPU, we can rule out backend/environment quirks. The likely issue is Arabic input processing.

    higgs_english_chunk300_unseeded_cuda.wav

    higgs_chinese_long_chunk300_unseeded_cuda.wav

    ok thank you @0xShug0 for explaining .. so i tested higgs model in english and the voice is not consistent and it didn't use reference voice but different voices as i said before

    ./build/linux-cpu-release/bin/audiocpp_cli
    --task tts
    --session-option "higgs_audio.mem_saver=true"
    --family higgs_audio_tts
    --model /home/xxxx/audio.cpp/build/linux-cpu-release/bin/models/Higgs-Audio-v3-TTS-4B-GGUF/higgs-audio-v3-tts-4b-q8_0.gguf
    --backend cpu
    --threads 10
    --text-chunk-size 300
    --text "$(cat /home/xxxx/audio.cpp/inputs/eng_test.txt)"
    --voice-ref /home/xxxx/audio.cpp/inputs/stoic_10seconds.wav
    --reference-text "$(cat /home/xxxx/audio.cpp/inputs/stoic_10seconds.txt)"
    --out /home/xxxx/audio.cpp/outputs/higgs_eng_1.wav
    --metrics
    --log
    eng_test.txt
    higgs_eng_1.wav
    higgs_eng_1 log.txt

  7. 0xShug0 commented on Sep 6, 2026

    @0xShug0
    0xShug0
    on Sep 6, 2026
    Owner

    @user6654126 I’ll focus on the consistency issue first. I ran your test case on CPU, and it produced a consistent voice. Are you using the prebuilts, or did you build from commit 49b0322?

    Another minor issue, caused by stale docs on my side: --session-option "higgs_audio.mem_saver=true" is not a valid option. It should be higgs_audio_tts.mem_saver. But in the test, I used your option exactly as written so it isn't relevent.

    higgs_eng_user_chunk300_mem_saver_cpu_e75c235.wav

  8. user6654126 commented on Sep 6, 2026

    @user6654126
    user6654126
    on Sep 6, 2026
    Author

    @user6654126 I’ll focus on the consistency issue first. I ran your test case on CPU, and it produced a consistent voice. Are you using the prebuilts, or did you build from commit 49b0322?

    Another minor issue, caused by stale docs on my side: --session-option "higgs_audio.mem_saver=true" is not a valid option. It should be higgs_audio_tts.mem_saver. But in the test, I used your option exactly as written so it isn't relevent.

    higgs_eng_user_chunk300_mem_saver_cpu_e75c235.wav

    @0xShug0 i build it with these commands for more info see last update logs.txt
    git pull origin main
    rm -rf build/
    ./scripts/build_linux.sh --backend cuda --cuda-arch native --native-model-manager --target audiocpp_cli --target audiocpp_server
    ./scripts/build_linux.sh --backend cpu --native-model-manager --target audiocpp_cli --target audiocpp_server

    last update logs.txt

    why is the voice is not consistent but keep changing especially at the end in your sample
    higgs_eng_user_chunk300_mem_saver_cpu_e75c235.wav

  9. 0xShug0 commented on Sep 6, 2026

    @0xShug0
    0xShug0
    on Sep 6, 2026
    Owner

    why is the voice is not consistent but keep changing especially at the end in your sample

    You are right. I missed the ending part. Now it looks like the CPU backend issue and I can debug on my machine.

    Linux/CUDA seems to be consisent (hopefully I didn't miss anything).

    higgs_eng_user_chunk300_mem_saver_cuda_e75c235.wav

    Update: Linux/Vulkan seems fine too.

    higgs_eng_user_chunk300_mem_saver_vulkan_e75c235.wav

    Metal also seems fine.
    higgs_eng_user_chunk300_mem_saver_metal_e75c235.wav

  10. 0xShug0
    added
    help wantedExtra attention is needed
    on Sep 6, 2026

  11. 0xShug0 commented on Sep 6, 2026

    @0xShug0
    0xShug0
    on Sep 6, 2026
    Owner

    For anyone interested in this issue:

    This may not be caused by missing cross-chunk reference reuse. Higgs already reuses the encoded reference audio codes across chunks, and it also has an AR reference-prefix KV reuse path when the prefix matches. Disabling prefix KV reuse on CPU did not fix the drift using the user provided test case (whille it passed a shorter text test).

    The real concern is narrower: with the same Higgs Audio TTS chunked voice-clone path, CUDA/Vulkan/Metal keep the reference voice reasonably consistent across chunks, while CPU can drift into noticeable different voices between chunks. That suggests a CPU-side instability or state/lifecycle issue when repeatedly running the same graph/code path across chunks.

    Failed attempts: CPU AR weight fp16 / precision tweak; Codec-only precision tweak; Using eager attention.

  12. user6654126 commented on Sep 6, 2026

    @user6654126
    user6654126
    on Sep 6, 2026
    Author

    @0xShug0 thank you for your help. Since the issue is clearly is with using cpu the higgs model doesn't use the reference voice provided in arabic and use random and non consistent voices . Is there is a way to use this model on my rtx 5060 8gb vram as i already tried even with chunk 300 and chunk 500 everytime i get oom error. By the way i noticed similar issue with using cpu in voxcpm2 in arabic which i will post about soon the voxcpm2 model use the reference voice and voice is consistent however when converting longer tts script more than 5 minutes i get EOC error.

  13. 0xShug0 commented on Sep 7, 2026

    @0xShug0
    0xShug0
    on Sep 7, 2026
    Owner

    Higgs: I made a correctness fix to a framework module, and it seems to have fixed the Arabic EOC issue on my side.

    Voxcpm2: This is more of an upstream model issue. Please check their issues. You can find discussions about noises, tag control and long-form quality there. The maintainer also mentioned that long text can indeed cause quality degradation, and the issue hasn’t been fully resolved yet.

    You can check the Python + CUDA output attached. It seems to start hallucinating very early and produces more noise over time.
    voxcpm_latest_direct_cuda_english_64k.mp3

    I also dded a new option --request-option voxcpm2.chunk_strategy=stateless as a workaround of the noises issue. However, it does not provide cross-chunk continuity, because each chunk is synthesized independently from the same original reference/prompt. The mode is intended for text without tags, since tags can make independently synthesized chunks diverge more from each other.

  14. user6654126 commented on Sep 7, 2026

    @user6654126
    user6654126
    on Sep 7, 2026
    Author

    Higgs: I made a correctness fix to a framework module, and it seems to have fixed the Arabic EOC issue on my side.

    Voxcpm2: This is more of an upstream model issue. Please check their issues. You can find discussions about noises, tag control and long-form quality there. The maintainer also mentioned that long text can indeed cause quality degradation, and the issue hasn’t been fully resolved yet.

    You can check the Python + CUDA output attached. It seems to start hallucinating very early and produces more noise over time. voxcpm_latest_direct_cuda_english_64k.mp3

    I also dded a new option --request-option voxcpm2.chunk_strategy=stateless as a workaround of the noises issue. However, it does not provide cross-chunk continuity, because each chunk is synthesized independently from the same original reference/prompt. The mode is intended for text without tags, since tags can make independently synthesized chunks diverge more from each other.

    @0xShug0 thanks . i am waiting for fix for Higgs model as it's currently is the best opensource tts model for arabic can you please tell me about the fix you made did you fix EOC issue for arabic on cpu and the issue of non consistent voices and the model ignore the reference voice and use random voice on its own and when it will be avaliable so i can update audio.cpp build and use it
    by the way i have crazy idea i don't know if it may fix that issue with higgs model but what if i use q5 or q6 instead of q8 on cuda can you please convert or tell me convert commands as i already tried that and the model didnt work

  15. 0xShug0 commented on Sep 7, 2026

    @0xShug0
    0xShug0
    on Sep 7, 2026
    Owner

    Arabic EOC issue. The fix has already been merged into the current main branch.

  16. 7 remaining items

  17. user6654126 commented on Sep 14, 2026

    @user6654126
    user6654126
    on Sep 14, 2026
    Author

    @0xShug0 thanks for your help i just tested the colab prebuilt https://github.com/0xShug0/audio.cpp/blob/main/Notebooks/colab_audio_cpp.ipynb
    but no model is working i used t4 gpu in colab and had there errors with s2 pro and higgs audio tts v3 in arabic and english with both q8 and bf16 versions. should i open an issue or what ? here is logs
    Session log
    Browser-side lifecycle and request events.
    Status
    ok
    Backend
    cuda
    Registered
    2
    Resident
    1
    5:57:05 PM Request failed: Failed to fetch
    5:56:31 PM Running Text to speech...
    5:55:02 PM Request failed: Failed to fetch
    5:55:02 PM Load failed: Failed to fetch
    5:54:50 PM Unloading Higgs Audio v3 TTS 4B (tts + clone, GGUF Q8) before loading Fish Audio S2 Pro (tts + clone/control tags, GGUF Q8).
    5:54:50 PM Loading Fish Audio S2 Pro (tts + clone/control tags, GGUF Q8)…
    5:54:50 PM Running Text to speech...
    5:54:03 PM Installer status unavailable: Failed to fetch
    5:54:00 PM Installer status unavailable: Failed to fetch
    5:52:59 PM Request failed: Failed to fetch
    5:52:59 PM Load failed: Failed to fetch
    5:52:59 PM Load failed: Failed to fetch
    5:52:53 PM Loading Higgs Audio v3 TTS 4B (tts + clone, GGUF Q8)…
    5:52:53 PM Running Text to speech...
    5:52:40 PM Unloading Higgs Audio v3 TTS 4B (tts + clone, GGUF Q8) before loading Higgs Audio v3 TTS 4B (tts + clone, GGUF Q8).
    5:52:40 PM Loading Higgs Audio v3 TTS 4B (tts + clone, GGUF Q8)…
    5:51:59 PM Installer status unavailable: Failed to fetch
    5:51:57 PM Installer status unavailable: Failed to fetch
    5:51:56 PM Fish Audio S2 Pro (tts + clone/control tags, GGUF Q8) GGUF Q8 installation is running in the background.
    5:51:14 PM Request failed: failed to allocate Higgs TTS AR prefill graph
    5:51:10 PM Running Text to speech...
    5:50:53 PM Request failed: failed to allocate Higgs TTS AR prefill graph
    5:50:48 PM Higgs Audio v3 TTS 4B (tts + clone, GGUF Q8) is resident and ready.
    5:50:47 PM Higgs Audio v3 TTS 4B (tts + clone, GGUF Q8) is resident and ready.
    5:50:36 PM Loading Higgs Audio v3 TTS 4B (tts + clone, GGUF Q8)…
    5:50:36 PM Running Text to speech...
    5:50:04 PM Unloading Higgs Audio v3 TTS 4B (tts + clone, GGUF Q8) before loading Higgs Audio v3 TTS 4B (tts + clone, GGUF Q8).
    5:50:04 PM Loading Higgs Audio v3 TTS 4B (tts + clone, GGUF Q8)…
    5:48:54 PM Installer status unavailable: Failed to fetch
    5:48:52 PM Installer status unavailable: Failed to fetch
    5:48:52 PM Request failed: Failed to fetch
    5:48:10 PM Running Text to speech...
    5:47:53 PM Installer status unavailable: Failed to fetch
    5:47:50 PM Request failed: Failed to fetch
    5:47:17 PM Running Text to speech...
    5:46:51 PM Installer status unavailable: Failed to fetch
    5:46:49 PM Installer status unavailable: Failed to fetch
    5:46:49 PM Request failed: Failed to fetch
    5:46:13 PM Running Text to speech...
    5:45:47 PM Installer status unavailable: Failed to fetch
    5:45:47 PM Request failed: Failed to fetch
    5:45:10 PM Running Text to speech...
    5:44:48 PM Installer status unavailable: Failed to fetch
    5:44:45 PM Installer status unavailable: Failed to fetch
    5:43:44 PM Request failed: Failed to fetch
    5:42:54 PM Higgs Audio v3 TTS 4B (tts + clone, GGUF Q8) is resident and ready.
    5:42:53 PM Loading Higgs Audio v3 TTS 4B (tts + clone, GGUF Q8)…
    5:42:53 PM Running Text to speech...
    5:42:23 PM Higgs Audio v3 TTS 4B (tts + clone, GGUF Q8) GGUF BF16 installation is running in the background.
    5:41:40 PM Package sizes unavailable: Failed to fetch
    5:41:40 PM Request failed: Failed to fetch
    5:41:16 PM Running Text to speech...
    5:40:18 PM Request failed: failed to allocate Higgs TTS AR prefill graph
    5:40:14 PM Running Text to speech...
    5:39:22 PM Request failed: failed to allocate Higgs TTS AR prefill graph
    5:39:17 PM Running Text to speech...
    5:38:56 PM Request failed: Enter text to generate.
    5:38:53 PM Running Text to speech...
    5:38:34 PM Request failed: Enter text to generate.
    5:38:34 PM Higgs Audio v3 TTS 4B (tts + clone, GGUF Q8) is resident and ready.
    5:38:33 PM Loading Higgs Audio v3 TTS 4B (tts + clone, GGUF Q8)…
    5:38:33 PM Running Text to speech...

  18. 0xShug0 commented on Sep 14, 2026

    @0xShug0
    0xShug0
    on Sep 14, 2026
    Owner

    No idea how colab works but "Open in Colab" works for me. This looks more like a colab issue than an audio.cpp issue, so it’s out of scope for the project.

    simplescreenrecorder-2026-09-14_11.40.41.mp4

  19. DrewThomasson commented on Sep 14, 2026

    @DrewThomasson
    DrewThomasson
    on Sep 14, 2026
    Contributor

    Colab is not broken

    Tested Higgs Audio TTS with voice cloning on a Colab T4 using the Q8_0 GGUF.

    the short Arabic sample generated successfully and appears to clone the reference voice.

    BryanCranstonRef.wav
    ShortArabicBCTest.wav

    BryanCranstonRef.mp4
    ShortArabicBCTest.mp4

    BUT it appears The long-form tests fail with failed to allocate Higgs TTS AR prefill graph

    `HTTPError: HTTP Error 500: Internal Server Error

The above exception was the direct cause of the following exception:

RuntimeError                              Traceback (most recent call last)
[/tmp/ipykernel_496/3364252775.py](https://localhost:8080/#) in <cell line: 0>()
    168         except urllib.error.HTTPError as error:
    169             error_body = error.read().decode("utf-8", errors="replace")
--> 170             raise RuntimeError(f"Generation failed (HTTP {error.code}): {error_body}") from error
    171 
    172         elapsed = time.monotonic() - started_at

RuntimeError: Generation failed (HTTP 500): {"error":{"message":"failed to allocate Higgs TTS AR prefill graph","type":"server_error"}}`

  20. DrewThomasson commented on Sep 14, 2026

    @DrewThomasson
    DrewThomasson
    on Sep 14, 2026
    Contributor

    idk it could be a T4-specific memory constraint

    But considering the failure occurs inside audio.cpp while allocating the Higgs AR prefill idk

  21. 0xShug0 commented on Sep 14, 2026

    @0xShug0
    0xShug0
    on Sep 14, 2026
    Owner

    @DrewThomasson Reducing the chunk size and max_tokens can help the model fit low VRAM

  22. DrewThomasson commented on Sep 14, 2026

    @DrewThomasson
    DrewThomasson
    on Sep 14, 2026
    Contributor

    I tried reducing the chunk size but it didn't seem to help

    Although all I tested was these changes so far

    text_chunk_size: 300 -> 150
max_tokens: 2048 -> 1024
  23. 0xShug0
    added a commit that references this issue on Sep 14, 2026

    Improve Higgs TTS token-limit error guidance (#522)

    582a63c

  24. 0xShug0 commented on Sep 14, 2026

    @0xShug0
    0xShug0
    on Sep 14, 2026
    Owner

    I tried reducing the chunk size but it didn't seem to help

    Although all I tested was these changes so far

    text_chunk_size: 300 -> 150
max_tokens: 2048 -> 1024

    Is the input English?

    Update: set chunk size = 128 (max_token default) the measured peak VRAM usage is about 5.6 gb so can fit 8GB GPU..

  25. DrewThomasson commented on Sep 14, 2026

    @DrewThomasson
    DrewThomasson
    on Sep 14, 2026
    Contributor

    I tried reducing the chunk size but it didn't seem to help

    Although all I tested was these changes so far

    text_chunk_size: 300 -> 150
max_tokens: 2048 -> 1024

    Is the input English?

    Update: set chunk size = 128 (max_token default) the measured peak VRAM usage is about 5.6 gb so can fit 8GB GPU..

    Was doing Arabic testing

  26. 0xShug0 commented on Sep 14, 2026

    @0xShug0
    0xShug0
    on Sep 14, 2026
    Owner

    @DrewThomasson
    There are two different issues here (1) The text chunker is currently limited in handling arbitrary Arabic input format. The text may not be chunked at all and cause OOM. The formation could help as a workaround #471 (comment) before we add a dedicated Arabic chunker.

    (2) Even with correct chunking, the model can hit the reach max_tokens before EOC error in some random chunks and abort. The official Python impl has the same issue, so it may be a model limitation. The difference is that the Python implementation hides this failure and returns a WAV that appears valid. E.g., Python hit the 2,048-token limit on just 123 Unicode codepoints of Arabic, including vowel marks, producing 81.64 seconds of audio. Tuning the temperature and increasing max_tokens (at the cost of higher VRAM usage) may help, but I haven't tested it myself.

  27. user6654126 commented on Sep 16, 2026

    @user6654126
    user6654126
    on Sep 16, 2026
    Author

    sorry for late replay (been busy) @0xShug0 thanks for your help however i didn't say there was an issue for using audio.cpp colab script. it work fine but the issue is with using some models in arabic such as higgs audio tts v3 as i'm about to say.
    just so to be clear i'm using colab free account t4 gpu and please to anyone who is going to replay. please reproduce using colab with both these two arabic scripts (with and without formation) and reference audio and reference audio text. also state if you used t4 gpu in free account or another gpu and send colab logs
    with default settings text_chunk_size: 1000 & tried 150 max_tokens: 1024 all failed
    @DrewThomasson thanks for your input however try to test with long samples +5 minutes for higgs tts v3 in arabic
    @0xShug0 from what you said in your last comment the issue are with the model so i will post the issue in their github page and let them know. however some said that there is no issue using higgs audio tts v3 original safetensors in comfyui so if you want to find out about that issue you could test on comfyui of course if you have 12 or 16 gb vram. because i can't i had oom errors with my 8gb vram . i may try to run this test on colab soon . also could this issue happen only with using gguf version.
    arabic.txt
    gemini-arabic.txt
    stoic_10seconds.txt
    stoic_10seconds.wav
    Session log
    Browser-side lifecycle and request events.
    Status
    ok
    Backend
    cuda
    Registered
    1
    Resident
    1
    4:59:10 PM Request failed: Failed to fetch
    4:58:23 PM Running Text to speech...
    4:58:08 PM Request failed: Failed to fetch
    4:57:29 PM Running Text to speech...
    4:57:06 PM Request failed: Failed to fetch
    4:56:10 PM Running Text to speech...
    4:56:05 PM Package sizes unavailable: Failed to fetch
    4:56:05 PM Request failed: Failed to fetch
    4:55:28 PM Higgs Audio v3 TTS 4B (tts + clone, GGUF Q8) is resident and ready.
    4:55:25 PM Loading Higgs Audio v3 TTS 4B (tts + clone, GGUF Q8)…
    4:55:25 PM Running Text to speech...
    4:55:01 PM Package sizes unavailable: Failed to fetch
    4:54:59 PM Request failed: Failed to fetch
    4:54:59 PM Load failed: Failed to fetch
    4:54:46 PM Loading Higgs Audio v3 TTS 4B (tts + clone, GGUF Q8)…
    4:54:46 PM Running Text to speech...
    4:54:03 PM Package sizes unavailable: Failed to fetch
    4:54:01 PM Package sizes unavailable: Failed to fetch
    4:53:02 PM Installer status unavailable: Failed to fetch
    4:53:00 PM Installer status unavailable: Failed to fetch
    4:49:57 PM Installer status unavailable: Failed to fetch
    4:49:55 PM Installer status unavailable: Failed to fetch
    4:48:55 PM Installer status unavailable: Failed to fetch
    4:47:54 PM Installer status unavailable: Failed to fetch
    4:46:50 PM Installer status unavailable: Failed to fetch
    4:44:48 PM Package sizes unavailable: Failed to fetch
    4:44:48 PM Installer status unavailable: Failed to fetch
    4:44:28 PM Higgs Audio v3 TTS 4B (tts + clone, GGUF Q8) GGUF Q8 installation is running in the background.

  28. 0xShug0 commented on Sep 16, 2026

    @0xShug0
    0xShug0
    on Sep 16, 2026
    Owner

    @user6654126 Yes please submit the issue to their official repo. In fact, there are open issues reporting the exact the same problem (not even limted to Arabic)!

    Recent: boson-ai/higgs-audio#188

    Also in their HF repo: https://huggingface.co/bosonai/higgs-tts-3-4b/discussions/18

    This one is probbaly about the old version of Higgs Audio TTS boson-ai/higgs-audio#163

    Unless we all hallucinate.

  29. user6654126 commented on Sep 27, 2026

    @user6654126
    user6654126
    on Sep 27, 2026
    Author

    quick update thanks @0xShug0 the issue is fixed i can use higgs on cuda and generate arabic tts with vocalization and voice is consistent however i didn't test on cpu for the previous issue.

    after updating to v0.8.2-audio8-perf-hotfix
    the higgs audio v3 tts is now working great on my 8gb vram and max vram usage was 6gb vram it works perfectly with arabic scripts with vocalization like this script below and the result is consistent voice (same as reference voice provided) however after many tests it has max of about 13kb script file about avg 7 minutes tts if i go any longer i get hit with this error :
    audiocpp_cli failed: Higgs TTS generation reached max_tokens (2048) before EOC for this text chunk; raise it with --max-tokens on the CLI or the "max_tokens" request option on the server, or lower --text-chunk-size / "text_chunk_size" so each chunk needs fewer generated frames
    i use these settings i tried increasing chunk to 1000 and max tokens to 4048 and still got EOC error message
    ./build/linux-cuda-release/bin/audiocpp_cli
    --task tts
    --family higgs_audio_tts
    --session-option "higgs_audio.mem_saver=true"
    --model /home/xxxx/audio.cpp/build/linux-cpu-release/bin/models/Higgs-Audio-v3-TTS-4B-GGUF/higgs-audio-v3-tts-4b-q8_0.gguf
    --backend cuda
    --text "$(cat /home/xxxx/audio.cpp/inputs/arabic-mini.txt)"
    --voice-ref /home/xxxx/audio.cpp/inputs/stoic_10seconds.wav
    --reference-text "$(cat /home/xxxx/audio.cpp/inputs/stoic_10seconds.txt)"
    --out /home/xxxx/audio.cpp/outputs/higgs_ara_1.wav
    --text-chunk-size 500
    --max-tokens 2048
    --metrics
    --log

    arabic-mini.txt
    stoic_10seconds.txt
    stoic_10seconds.wav
    higgs log.txt
    higgs_ara_1.wav

  30. 0xShug0
    closed this as completedon Sep 27, 2026

  31. 0xShug0 commented on Sep 27, 2026

    @0xShug0
    0xShug0
    on Sep 27, 2026
    Owner

    @DrewThomasson There are two different issues here (1) The text chunker is currently limited in handling arbitrary Arabic input format. The text may not be chunked at all and cause OOM. The formation could help as a workaround #471 (comment) before we add a dedicated Arabic chunker.

    (2) Even with correct chunking, the model can hit the reach max_tokens before EOC error in some random chunks and abort. The official Python impl has the same issue, so it may be a model limitation. The difference is that the Python implementation hides this failure and returns a WAV that appears valid. E.g., Python hit the 2,048-token limit on just 123 Unicode codepoints of Arabic, including vowel marks, producing 81.64 seconds of audio. Tuning the temperature and increasing max_tokens (at the cost of higher VRAM usage) may help, but I haven't tested it myself.

    Please check my previous comments and report the issue to the higgs tts repo.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    help wantedExtra attention is needed

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions