--- license: apache-2.0 ---

FireRedAudio

Official PyTorch code for
FireRedAudio: A General-Purpose Audio Language Model with Decoupled Continuous Representations for Understanding and Generation

technical report version HF-model Apache-2.0
## Overview > **One model to listen, understand, reason, speak, and edit.** **FireRedAudio** is a general-purpose audio language model built on a shared **9B-parameter LLM** with **decoupled continuous representations**: an Audio Encoder handles understanding, while a RedAE pathway handles generation. A single model supports **ASR, audio understanding, zero-shot TTS, instruct TTS, semantic/acoustic speech editing, and accurate temporal grounding over recordings up to one hour long**.
FireRedAudio Logo
## Highlights ✨ * 🧩 **Purpose-built representations, one shared backbone** — The Audio Encoder pathway serves understanding, while the RedAE-Patch pathway serves speech generation. Their representations remain decoupled but share the same language and reasoning backbone. To the best of our knowledge, this is the first publicly disclosed design of its kind in a unified audio-language model. * 📊 **One model, a full audio stack** — FireRedAudio spans ASR, broad and fine-grained audio understanding, zero-shot TTS, Instruct TTS, and free-form speech editing, achieving competitive or leading results across MMAU, MMSU, Seed-TTS-Eval, InstructTTSEval, and Ming-Freeform-Audio-Edit. * 🎙️ **Create and edit speech with natural language** — Clone a voice from a reference clip, design a voice from a description, or edit what was said and how it sounds through one continuous-latent generation pathway. * ⏱️ **Go from minutes to hour-long recordings** — Understand recordings up to one hour with precise time-to-content alignment. Organize audio into timestamped structures, produce grounded summaries, retrieve content by time (or time by content), and reason over evidence distributed across the recording. ## News - [2026.08.21] We release the **FireRedAudio** code and model. ## Contents - [Quick Start](#quick-start-) - [Performance](#performance) - [Limitations](#limitations) - [Usage Disclaimer](#usage-disclaimer) - [Citation](#citation) - [Acknowledgements](#acknowledgements) - [License](#license) ## Quick Start 🚀 ### Installation Requires **Python 3.10** and [uv](https://docs.astral.sh/uv/). System prerequisites: a **CUDA toolkit** (GPU inference + compiling the causal-conv1d / flash-attn kernels) and **ffmpeg** (audio decoding via torchaudio/torchcodec). Wheels target **CUDA 12.8 (`cu128`)** by default. For a different CUDA version, change the `pytorch-cu128` index URL in [`pyproject.toml`](pyproject.toml) (e.g. `cu126` / `cu129`) and re-run `uv sync`. ```sh uv sync --extra accel --extra accel-build --group tools # + compile causal-conv1d, flash-attn ``` ### Model Download Download the pretrained model from Hugging Face with the `hf` CLI: ```sh hf download FireRedTeam/FireRedAudio --local-dir pretrained_models/ ``` ### Python API ```python import torch import torchaudio from inference import FireRedAudioInference # Init the model. Understanding tasks need only --model; generation tasks # additionally need the RedAE decoder weights. engine = FireRedAudioInference( model_path="pretrained_models/FireRedAudio", vae_decoder_path="pretrained_models/RedAE_decoder/model.pt", # only for tts / edit / voice_design device="cuda:0", ) # ---- 1) Speech recognition (ASR) ----------------------------------------- res = engine.understand("assets/examples/asr_zh_fleurs.wav", "Transcribe speech to text.", task="asr") print(res.answer) # ---- 2) Audio understanding (with optional chain-of-thought) -------------- res = engine.understand( "assets/examples/two_speakers.wav", "这个音频中有几个说话人", task="understand", enable_thinking=True, max_new_tokens=10240, ) print("CoT:") print(res.reasoning) # CoT reasoning, or None print("Answer") print(res.answer) # ---- 3) Zero-shot TTS (ICL voice cloning) --------------------------------- res = engine.tts( prompt_text="同时，他强调微调要科学有序。", prompt_audio="assets/examples/tts_zh_prompt.wav", target_text="安徽淮南秦师傅发现，停在小区的爱车右前驾驶窗玻璃被砸。", language="zh", ) torchaudio.save("tts.wav", res.audio.cpu(), sample_rate=24000) # ---- 4) Speech editing ----------------------------------------------------- # semantic: rewrite / substitute / insert / delete content. The model first writes # <|sot|>{rewritten text}<|eot|> then renders the audio. res = engine.edit("assets/examples/edit_semantic_zh_ref.wav", "delete '比普通的茶叶要'", edit_type="semantic" print(res.text) torchaudio.save("edit_semantic.wav", res.audio.cpu(), sample_rate=24000) # acoustic: change pitch / speed / volume. The instruction must follow the exact # templates below (the model is trained on these, not free-form phrasing): # pitch -> "shift the pitch by N step(s)" N in {-6, ..., -1, 1, ..., +6} # speed -> "adjust the speed to X" X in [0.5, 2.0], step 0.1 # volume -> "adjust the volume to X" X in [0.3, 2.0], step 0.1 res = engine.edit("assets/examples/edit_acoustic_zh_ref.wav", "shift the pitch by 3 steps", edit_type="acoustic") torchaudio.save("edit_acoustic.wav", res.audio.cpu(), sample_rate=24000) # ---- 5) Voice design (synthesis from a timbre description) ----------------- res = engine.voice_design( instruction="以女性高音区的清亮音色,表现出青年阶段的特质,音量略强,语速适中稍快,语调带有解释意味和急切的情感流露,确保语音流畅自然。", text="是我请他来的，可他什么也不知道，他来只是想打听一下，你们厂是不是有旧锅炉？", ) torchaudio.save("voice_design.wav", res.audio.cpu(), sample_rate=24000) ``` ### Command Line Inference ```sh # speech recognition uv run inference.py --task asr --model pretrained_models/FireRedAudio --audio assets/examples/asr_zh_fleurs.wav # audio understanding and QA; several --audio for e.g. speaker verification, # --enable-thinking to let the model reason first uv run inference.py --task understand --model pretrained_models/FireRedAudio --audio assets/examples/two_speakers.wav \ --prompt "这个音频中有几个说话人" --enable-thinking --max-new-tokens 4096 # ICL voice cloning from a reference audio and its transcript uv run inference.py --task tts --model pretrained_models/FireRedAudio --vae-decoder pretrained_models/RedAE_decoder/model.pt \ --prompt-audio assets/examples/tts_zh_prompt.wav --prompt-text "同时，他强调微调要科学有序。" \ --target-text "安徽淮南秦师傅发现，停在小区的爱车右前驾驶窗玻璃被砸。" --language zh --output tts.wav # speech editing; semantic rewrites content, acoustic changes pitch / speed / volume. uv run inference.py --task edit --model pretrained_models/FireRedAudio --vae-decoder pretrained_models/RedAE_decoder/model.pt \ --audio assets/examples/edit_semantic_zh_ref.wav --instruction "delete '比普通的茶叶要'" --edit-type semantic \ --output edit_semantic.wav # acoustic: instructions must use the exact trained templates, e.g. # "shift the pitch by N step(s)" in -6..6 steps (pitch) # "adjust the speed to X" in [0.5, 2.0], step .1 (speed) # "adjust the volume to X" in [0.3, 2.0], step .1 (volume) uv run inference.py --task edit --model pretrained_models/FireRedAudio --vae-decoder pretrained_models/RedAE_decoder/model.pt \ --audio assets/examples/edit_acoustic_zh_ref.wav --instruction "shift the pitch by 3 steps" --edit-type acoustic \ --output edit_acoustic.wav # voice design uv run inference.py --task voice_design --model pretrained_models/FireRedAudio --vae-decoder pretrained_models/RedAE_decoder/model.pt \ --instruction "以女性高音区的清亮音色,表现出青年阶段的特质,音量略强,语速适中稍快,语调带有解释意味和急切的情感流露,确保语音流畅自然。" \ --text "是我请他来的，可他什么也不知道，他来只是想打听一下，你们厂是不是有旧锅炉？" --output voice_design.wav ``` ## Performance ### Audio Understanding
Model
MMAU test-mini
MMAU test
MMSU
Step-Audio-R1.177.7–75.9
Step-Audio 278.0––
MiMo-Audio-7B-Instruct74.9–61.7
Kimi-Audio65.2––
LongCat-Next76.4––
Qwen3-Omni-30B-A3B-Instruct77.5–69.0
Gemini 3.1 Pro80.7*78.8*82.7*
Qwen3.5-Omni-Plus81.4*79.9*80.7*
FireRedAudio81.780.983.3
* *Results marked with * are obtained from our own evaluation.* ### ASR
Model
AISHELL‑1
AISHELL‑2
test‑ios
WenetSpeech
Net ⁠|⁠ Meeting
LibriSpeech
clean ⁠|⁠ other
FLEURS
en ⁠|⁠ zh
FLEURS‑102
avg
KeSpeech
Opencpop
Step‑Audio 20.632.104.67 ⁠|⁠ 4.751.17 ⁠|⁠ 2.423.03 ⁠|⁠ 2.68–3.63–
MiMo‑Audio‑7B‑Instruct1.65––3.50 ⁠|⁠ –––––
Ming‑UniAudio‑16B‑A3B–2.84–1.62 ⁠|⁠ –––––
Kimi‑Audio0.602.565.37 ⁠|⁠ 6.281.28 ⁠|⁠ 2.424.44 ⁠|⁠ 2.69–––
LongCat‑Next1.472.825.98 ⁠|⁠ 8.191.63 ⁠|⁠ 3.425.24 ⁠|⁠ 3.24–––
Qwen3‑Omni‑30B‑A3B‑Instruct––4.69 ⁠|⁠ 5.891.22 ⁠|⁠ 2.482.72 ⁠|⁠ 2.20––1.54
Gemini 3.1 Pro3.66*7.10*11.53 ⁠|⁠ 14.213.36 ⁠|⁠ 4.412.97* ⁠|⁠ 4.28*18.23*23.676.83
Qwen3.5‑Omni‑Plus0.82*2.26*4.30 ⁠|⁠ 5.841.11 ⁠|⁠ 2.233.33* ⁠|⁠ 2.46*23.66*3.461.49
FireRedAudio0.712.595.18 ⁠|⁠ 5.330.69 ⁠|⁠ 2.902.61 ⁠|⁠ 3.2115.024.941.63
* *Results marked with * are obtained from our own evaluation.* ### Zero-Shot TTS
Model
Seed-ZH
CER↓ ⁠|⁠ SIM↑
Seed-EN
WER↓ ⁠|⁠ SIM↑
Avg.
CER/WER↓ ⁠|⁠ SIM↑
Seed-TTS1.12 ⁠|⁠ 0.802.25 ⁠|⁠ 0.761.69 ⁠|⁠ 0.78
FireRedTTS1.51 ⁠|⁠ 0.653.82 ⁠|⁠ 0.532.67 ⁠|⁠ 0.59
FireRedTTS-21.14 ⁠|⁠ 0.7361.95 ⁠|⁠ 0.651.55 ⁠|⁠ 0.69
DiTAR (1B)1.02 ⁠|⁠ 0.7531.69 ⁠|⁠ 0.741.36 ⁠|⁠ 0.75
F5-TTS1.56 ⁠|⁠ 0.741.83 ⁠|⁠ 0.651.70 ⁠|⁠ 0.70
CosyVoice 21.45 ⁠|⁠ 0.752.57 ⁠|⁠ 0.652.01 ⁠|⁠ 0.70
CosyVoice 3-1.5B1.12 ⁠|⁠ 0.782.21 ⁠|⁠ 0.721.67 ⁠|⁠ 0.75
MiMo-Audio-7B-Instruct1.96 ⁠|⁠ –5.37 ⁠|⁠ –3.67 ⁠|⁠ –
Qwen2.5-Omni-7B (RL)1.42 ⁠|⁠ 0.752.33 ⁠|⁠ 0.641.88 ⁠|⁠ 0.70
Qwen3-Omni-30B-A3B-Instruct1.07 ⁠|⁠ –1.39 ⁠|⁠ –1.23 ⁠|⁠ –
Ming-UniAudio-16B-A3B0.95 ⁠|⁠ 0.701.85 ⁠|⁠ 0.581.40 ⁠|⁠ 0.64
FireRedAudio0.86 ⁠|⁠ 0.741.72 ⁠|⁠ 0.681.29 ⁠|⁠ 0.71
### Instruct TTS
Model
ZH
APS↑ ⁠|⁠ DSD↑ ⁠|⁠ RP↑
EN
APS↑ ⁠|⁠ DSD↑ ⁠|⁠ RP↑
VoiceSculptor-VD74.6 ⁠|⁠ 63.5 ⁠|⁠ 62.0– ⁠|⁠ – ⁠|⁠ –
MOSS-VoiceGenerator71.6 ⁠|⁠ 72.5 ⁠|⁠ 61.358.8 ⁠|⁠ 71.8 ⁠|⁠ 61.6
Ming-Omni-TTS-16B84.6 ⁠|⁠ 70.7 ⁠|⁠ 56.0– ⁠|⁠ – ⁠|⁠ –
Qwen3-TTS-VD83.7 ⁠|⁠ 81.7 ⁠|⁠ 65.876.4 ⁠|⁠ 81.4 ⁠|⁠ 64.2
FireRedAudio87.1 ⁠|⁠ 82.3 ⁠|⁠ 69.980.3 ⁠|⁠ 84.3 ⁠|⁠ 71.6
* *These results are obtained from our own evaluation.* ### Semantic Editing
Task
Setting
Metric
Ming-UniAudio-Edit
zh | en
FireRedAudio
zh | en
Deletion basic WER (%)↓ 11.89 | 14.85 11.09 | 13.20
SIM↑0.78 | 0.760.78 | 0.79
ACC (%)↑100 | 82.2299.42 | 96.67
no-edit WER (%)↓11.49 | 24.2610.75 | 23.80
open WER (%)↓ 22.92 | 27.60 10.86 | 16.97
SIM↑0.81 | 0.740.81 | 0.80
ACC (%)↑82.92 | 85.0088.97 | 86.50
no-edit WER (%)↓17.50 | 35.218.11 | 25.69
Insertion basic WER (%)↓ 3.42 | 6.63 3.05 | 5.42
SIM↑0.83 | 0.790.83 | 0.84
ACC (%)↑80.00 | 71.4383.53 | 85.09
no-edit WER (%)↓3.52 | 17.703.34 | 16.88
open WER (%)↓ 3.89 | 7.59 2.67 | 6.49
SIM↑0.83 | 0.790.83 | 0.84
ACC (%)↑79.31 | 62.3186.21 | 70.85
no-edit WER (%)↓4.10 | 18.842.94 | 17.56
Substitution basic WER (%)↓ 4.52 | 8.99 2.60 | 4.82
SIM↑0.82 | 0.780.84 | 0.81
ACC (%)↑78.62 | 59.7887.42 | 74.86
no-edit WER (%)↓4.63 | 19.282.79 | 16.92
open WER (%)↓ 4.56 | 7.64 2.21 | 3.83
SIM↑0.83 | 0.770.84 | 0.81
ACC (%)↑76.62 | 65.6291.38 | 76.56
no-edit WER (%)↓4.75 | 18.392.46 | 15.69
### Acoustic Editing
Task
Metric
Ming-UniAudio-Edit
zh | en
FireRedAudio
zh | en
Speed Alteration WER (%)↓ 5.88 | 17.53 1.86 | 4.71
SIM↑0.66 | 0.570.79 | 0.71
RDE (%)↓6.36 | 5.922.60 | 4.02
Pitch Alteration WER (%)↓ 7.45 | 13.37 1.60 | 2.89
SIM↑0.36 | 0.240.51 | 0.44
Volume Alteration WER (%)↓ 1.71 | 1.35 1.60 | 1.12
SIM↑0.86 | 0.800.94 | 0.93
RAE (%)↓14.9 | 11.72.64 | 3.75
## Limitations - **Everything but ASR is Chinese/English only.** Speech generation (`tts` / `edit` / `voice_design`) and audio understanding are limited to Chinese and English — `tts` selects the language via `--language zh` / `en`. ASR is the only task that supports more languages. - **Zero-shot TTS is not deterministic by default.** The flow-matching decoder samples random noise, so output varies run to run; pass a fixed seed (`set_seed(...)` in the API, `--seed` on the CLI) for reproducibility, and note that quality can differ across seeds. - **Long-form input is supported up to about one hour.** Beyond that the model is untested and time-to-content alignment may degrade. ## Usage Disclaimer - The project incorporates zero-shot voice cloning functionality; Please note that this capability is intended **solely for academic research purposes**. - **DO NOT** use this model for **ANY illegal activities**❗️❗️ - The developers assume no liability for any misuse of this model. - If you identify any instances of **abuse**, **misuse**, or **fraudulent** activities related to this project, **please report them to our team immediately.** ## Citation ```bib @article{fireredaudio, title = {FireRedAudio: A General-Purpose Audio Language Model with Decoupled Continuous Representations for Understanding and Generation}, author = {FireRed Team}, journal = {arXiv preprint}, year = {2026}, } ``` ## Acknowledgements - [Qwen3.5](https://github.com/QwenLM/Qwen3.8) for the language model foundation - [Whisper-large-v3](https://huggingface.co/openai/whisper-large-v3) for the Audio Encoder initialization - [x-transformers](https://github.com/lucidrains/x-transformers) for RotaryEmbedding - [vocos](https://github.com/gemelo-ai/vocos/tree/main) for ISTFT implementation ## License Released under the [Apache-2.0](LICENSE) license.