From a6b5c7231c831195b8a2ea17a5bbb1c5a334d3e2 Mon Sep 17 00:00:00 2001
From: audio2text dev <dev@audio2text.local>
Date: Mon, 6 Jul 2026 22:28:37 +0800
Subject: [PATCH] =?UTF-8?q?perf:=20ASR=20GPU=20=E5=88=A9=E7=94=A8=E7=8E=87?=
 =?UTF-8?q?=E4=BC=98=E5=8C=96=20=E2=80=94=20batch=5Fsize=2016=E2=86=9232?=
 =?UTF-8?q?=20+=20beam=5Fsize=205=E2=86=922?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

faster-whisper 的 GPU 利用率呈尖刺波（峰=批量解码满载，谷=CPU 提取 Mel
特征 + 处理结果时 GPU 空闲），平均利用率低。瓶颈不在算力而在 CPU/GPU
未重叠。

- batch_size 16→32：拉长单次 GPU 解码时间，相对掩盖 CPU 特征提取间隙，
  尖刺变宽变平，平均利用率上升。turbo FP16 仅 ~1.6GB，3090 24G 充裕。
- beam_size 5→2：turbo 模型鲁棒，候选数 5→2 大幅减少解码步数，让 GPU
  峰更密、间隙更短。保留 1 个候选做歧义发音保险，质量损失小。
- beam_size 从硬编码提到 config 可调，CPU/CPU 模板/GPU/示例 四份配置对齐
- /health 增加 asr_beam_size，模型加载日志同步输出 batch+beam

word_timestamps 保留 True：segmenter 强依赖词级时间戳做精确断句，
关闭会触发匀速估算退化路径，得不偿失。
---
 app/config.py                 | 1 +
 app/main.py                   | 1 +
 app/services/asr_service.py   | 2 +-
 app/services/model_manager.py | 3 ++-
 config.cpu.yaml               | 1 +
 config.example.yaml           | 3 ++-
 config.gpu.yaml               | 3 ++-
 7 files changed, 10 insertions(+), 4 deletions(-)

diff --git a/app/config.py b/app/config.py
index c632a10..3369ed3 100644
--- a/app/config.py
+++ b/app/config.py
@@ -48,6 +48,7 @@ class AsrConfig(BaseModel):
     word_timestamps: bool = True
     vad_filter: bool = True
     batch_size: int = 8          # BatchedInferencePipeline 的音频块批大小；GPU 建议 16
+    beam_size: int = 5           # beam search 宽度；GPU turbo 可降到 2 加速（候选数↓ 解码步数↓），质量损失小
 
 
 class TranslationConfig(BaseModel):
diff --git a/app/main.py b/app/main.py
index 6e32ded..98eee80 100644
--- a/app/main.py
+++ b/app/main.py
@@ -171,6 +171,7 @@ def create_app() -> FastAPI:
         info["asr_model"] = s.asr.model
         info["asr_compute_type"] = s.asr.compute_type
         info["asr_batch_size"] = s.asr.batch_size
+        info["asr_beam_size"] = s.asr.beam_size
         info["asr_language"] = s.asr.language
         info["translation_device"] = s.translation.device
         info["translation_model"] = s.translation.model
diff --git a/app/services/asr_service.py b/app/services/asr_service.py
index d526f50..9d2e40b 100644
--- a/app/services/asr_service.py
+++ b/app/services/asr_service.py
@@ -37,7 +37,7 @@ def transcribe(wav_path: Path) -> list[Segment]:
         language=s.language,
         word_timestamps=s.word_timestamps,
         vad_filter=s.vad_filter,
-        beam_size=5,
+        beam_size=s.beam_size,        # beam search 宽度（config 可调）；GPU turbo 建议降到 2 加速
         batch_size=s.batch_size,       # 批量解码：多音频块一次性送 GPU
         without_timestamps=False,      # BatchedInferencePipeline 默认 True，需显式关闭以生成段级时间戳
     )
diff --git a/app/services/model_manager.py b/app/services/model_manager.py
index 8b1cc49..27c418e 100644
--- a/app/services/model_manager.py
+++ b/app/services/model_manager.py
@@ -65,7 +65,8 @@ class ModelManager:
             )
             self._asr = BatchedInferencePipeline(model=whisper)
             self._current = "asr"
-            logger.info("ASR 模型已就绪（batched, batch_size=%d）。", s.batch_size)
+            logger.info("ASR 模型已就绪（batched, batch_size=%d, beam_size=%d）。",
+                        s.batch_size, s.beam_size)
             return self._asr
 
     def unload_asr(self) -> None:
diff --git a/config.cpu.yaml b/config.cpu.yaml
index 3547ad2..63ba5bc 100644
--- a/config.cpu.yaml
+++ b/config.cpu.yaml
@@ -31,6 +31,7 @@ asr:
   word_timestamps: true
   vad_filter: true
   batch_size: 8             # CPU 无 GPU 并行收益，保持小批
+  beam_size: 5              # tiny.en CPU 质量优先，保持默认 beam（无加速诉求）
 
 translation:
   model: Helsinki-NLP/opus-mt-en-zh   # 最轻量英译中（~300MB；NLLB-600M 需 ~2.4GB，2GB 机 OOM）
diff --git a/config.example.yaml b/config.example.yaml
index f8cb167..7d7af55 100644
--- a/config.example.yaml
+++ b/config.example.yaml
@@ -32,7 +32,8 @@ asr:
   language: en                         # 仅英语
   word_timestamps: true                # 词级时间戳：让断句精确而非纯匀速估算
   vad_filter: true                     # 过滤静音段，提升质量与速度
-  batch_size: 8                        # CPU: 8 | GPU: 16（BatchedInferencePipeline 批量解码音频块）
+  batch_size: 8                        # CPU: 8 | GPU: 32（BatchedInferencePipeline 批量解码音频块，拉长单次 GPU 解码掩盖 CPU 特征提取间隙）
+  beam_size: 5                         # CPU: 5（默认）| GPU: 2（turbo 鲁棒可降，候选数↓解码步数↓，加速明显质量损失小）
 
 translation:
   # CPU dev：Helsinki-NLP/opus-mt-en-zh（~300MB，2GB 内存可跑）
diff --git a/config.gpu.yaml b/config.gpu.yaml
index e793899..9e414f1 100644
--- a/config.gpu.yaml
+++ b/config.gpu.yaml
@@ -29,7 +29,8 @@ asr:
   language: en
   word_timestamps: true
   vad_filter: true
-  batch_size: 16             # BatchedInferencePipeline：每批解码 16 个 30s 音频块，填充 GPU
+  batch_size: 32             # BatchedInferencePipeline：每批解码 32 个 30s 音频块，拉长单次 GPU 解码掩盖 CPU 特征提取间隙
+  beam_size: 2               # turbo 鲁棒，beam=2 留一个候选做保险（vs beam=5 候选数↓ 解码步数↓，加速明显质量损失小）
 
 translation:
   model: facebook/nllb-200-distilled-1.3B   # 质量最好