语音识别案例

本案例介绍如何定义Vectorized Scalar UDF来进行语音识别、如何定义UDAF来进行聚合统计和可视化。

1. 从fabric_data中引入高阶类型（图片、音频、视频）：

   from fabric_data.multimodal.types.image import Image, display_image
   from fabric_data.multimodal.types.audio import Audio
   from fabric_data.multimodal.types.vectorized import PandasVector

2. 定义Class Vectorized Scalar UDF来进行语音识别，这里使用到外部包transformers、torch：

   class SpeechRecognitionUDF:
       def __init__(self, model_path: str) -> None:
           import os
           from transformers import pipeline
           import torch
   
           model_dir = os.path.abspath(model_path)
           self._asr = pipeline(
               "automatic-speech-recognition",
               model=model_dir,
               device=0 if torch.cuda.is_available() else -1,
           )
   
       def __call__(self, audios: PandasVector[Audio]) -> PandasVector[str]:
           return PandasVector[str](audios.map(
               lambda audio: self._asr(audio.sound_file.read(dtype="float32"))["text"]
           ))

3. 定义UDAF来进行TF-IDF词频统计和词云图片生成，这里使用到外部包wordcloud、matplotlib：

   class SpeakerWordsUDAF:
       def __init__(self, top_n: int = 40):
           from typing import List
           self._texts: List[str] = []
           self._top_n = top_n
           self._chapter_id = None
   
           # tiny built-in stopword list (extend as needed)
           self._stopwords = {
               "the", "a", "an", "and", "or", "but",
               "also", "about", "this", "that", "these", "those",
               "his", "her", "its", "their", "our", "your",
               "of", "to", "in", "on", "for", "at", "by", "with",
               "from", "as", "is", "are", "was", "were", "be", "been",
               "it", "he", "she", "they", "we", "you", "i",
               "not", "do", "did", "does", "had", "have", "has",
               "how", "what", "when", "where", "which", "who", "whom",
               "would", "could", "should", "will", "can", "may",
           }
   
       @property
       def aggregate_state(self):
           return {
               "texts": self._texts,
               "chapter_id": self._chapter_id,
           }
   
       def accumulate(self, chapter_id: int, hyp_text: str):
           if hyp_text:
               self._texts.append(hyp_text)
           if chapter_id:
               self._chapter_id = chapter_id
   
       def merge(self, other_state):
           if other_texts := other_state.get("texts"):
               self._texts.extend(other_texts)
           if other_chapterid := other_state.get("chapter_id"):
               if self._chapter_id:
                   # for one SpeakerWordsUDAF instance, it should process texts with same chapter_id
                   assert self._chapter_id == other_chapterid
               else:
                   # CN get chapter_id from DN
                   self._chapter_id = other_chapterid
   
       def _tokenize(self, text: str):
           # simple whitespace tokenizer + basic cleaning
           tokens = []
           for t in text.split():
               t = t.strip().lower()
               # drop short tokens and non-alpha tokens
               if len(t) < 3:
                   continue
               if not t.isalpha():
                   continue
               if t in self._stopwords:
                   continue
               tokens.append(t)
           return tokens
   
       def _generate_wordcloud(self, frequencies: dict) -> Image:
           from io import BytesIO
           import matplotlib.pyplot as plt
           from wordcloud import WordCloud
           try:
               # Create figure
               fig, ax = plt.subplots(figsize=(10, 5))
   
               # Generate WordCloud
               wc = WordCloud(width=800, height=400, background_color="white") \
                   .generate_from_frequencies(frequencies)
   
               # Draw
               ax.imshow(wc, interpolation='bilinear')
               ax.axis("off")
               ax.set_title(f"WordCloud for Chapter {self._chapter_id}", fontsize=18, pad=20)
   
               # Save to in-memory bytes
               buf = BytesIO()
               fig.savefig(buf, format="png", bbox_inches="tight")
               buf.seek(0)
   
               return Image(data=buf.getvalue(), format="PNG")
           except Exception as e:
               # You may log or re-raise here
               raise
           finally:
               # Cleanup: ALWAYS close the figure
               if fig is not None:
                   plt.close(fig)
   
       def finish(self) -> Image:
           import math
           if not self._texts:
               return None
   
           # 1) tokenize all docs
           docs = [self._tokenize(txt) for txt in self._texts]
           num_docs = len(docs)
   
           # 2) document frequency
           df = {}
           for tokens in docs:
               for tok in set(tokens):
                   df[tok] = df.get(tok, 0) + 1
   
           # 3) tf-idf per doc → aggregate
           tfidf_sums = {}
           tfidf_counts = {}
   
           for tokens in docs:
               if not tokens:
                   continue
               # term freq in this doc
               tf = {}
               for tok in tokens:
                   tf[tok] = tf.get(tok, 0) + 1
   
               doc_len = len(tokens)
               for tok, f in tf.items():
                   tf_val = f / doc_len
                   # idf: log(N/df) + 1
                   idf_val = math.log(num_docs / df[tok]) + 1.0
                   tfidf = tf_val * idf_val
   
                   tfidf_sums[tok] = tfidf_sums.get(tok, 0.0) + tfidf
                   tfidf_counts[tok] = tfidf_counts.get(tok, 0) + 1
   
           # 4) average tf-idf per term
           avg_tfidf = {
               tok: (tfidf_sums[tok] / tfidf_counts[tok])
               for tok in tfidf_sums
           }
   
           # 5) sort & top-N
           sorted_terms = sorted(
               avg_tfidf.items(),
               key=lambda kv: kv[1],
               reverse=True
           )[: self._top_n]
   
           out = {word: float(score) for word, score in sorted_terms}
           return self._generate_wordcloud(out)

4. 建立远程连接，这里用到外部包huawei-fabricsql-connectorapi：

   from contextlib import contextmanager
   from fabric_data.multimodal import ai_lake
   import os
   
   @contextmanager
   def create_connect():
       conn = ai_lake.connect(
           fabric_endpoint="172.16.31.57:60001",
           fabric_endpoint_id="8a708bbf-f862-4c53-9622-4c59b44524a8",
           fabric_workspace_id="ca319048-b07c-498c-97df-8f20b8ce0a52",
           lf_catalog_name="mini_kernel2",
           lf_instance_id="e662709e-6223-44a8-b254-d84cf71842de",
           access_key=os.getenv("access_key"),
           secret_key=os.getenv("secret_key"),
           default_database="multimodal_lake",
           use_single_cn_mode=True,
       )
       print(f"Establish ai_lake.connect: {conn}")
       try:
           yield conn
       finally:
           # release resource
           conn.close()

5. 准备Parquet表：

   with create_connect() as conn:
       from fabric_data.ibis.backends.sql.datatype import advance_dtype
       conn.drop_table("librispeech_asr", database="multimodal_lake")
       table = conn.create_table(
           "librispeech_asr",
           schema={
               "file": str,
               "audio": advance_dtype(Audio),
               "text": str,
               "speaker_id": int,
               "chapter_id": int,
               "id": str,
           },
           table_format="parquet",
           external=True,
           if_not_exists=True,
           location="obs://mini-kernel/mini_kernel2/multimodal-test/librispeech_asr",
           database="multimodal_lake",
       )
       print(table)

6. 注册Scalar UDF，UDAF：

   with create_connect() as conn:
       conn.set_function_staging_workspace(
           obs_server=os.getenv("obs_server"),
           obs_bucket_name="mini-kernel",
           obs_directory_base="mini_kernel2/multimodal-test",
           access_key=os.getenv("access_key"),
           secret_key=os.getenv("secret_key"),
       )
       print("finish set_function_staging_workspace")
   
       conn.delete_function("SpeechRecognitionUDF", database="multimodal_lake")
       conn.create_scalar_function(
           SpeechRecognitionUDF,
           database="multimodal_lake",
           packages=["transformers", "torch"],
           imports=["./facebook_wav2vec2_base_960h"]
       )
   
       conn.delete_function("SpeakerWordsUDAF", database="multimodal_lake")
       conn.create_agg_function(
           SpeakerWordsUDAF,
           database="multimodal_lake",
           packages=["wordcloud", "matplotlib"]
       )

7. 以Fabric Data API形式使用Scalar UDF, UDAF，得到最后的分组统计词云图片：

   with create_connect() as conn:
       # load dataset
       ds = conn.load_dataset("librispeech_asr", database="multimodal_lake")
       ds.show(limit=1)
       ds = ds.limit(1)
   
       # use Scalar UDF for ASR
       udf = conn.get_function("SpeechRecognitionUDF", database="multimodal_lake")
   
       ds = ds.map_batches(
           fn=udf,
           on=[ds.audio],
           as_col="hyp_text",
           num_dpus=0.5,
           concurrency=1,
           model_path="./fabric_data/examples/facebook_wav2vec2_base_960h",
       )
       ds.show(limit=1)
   
       # use UDAF for summary
       udaf = conn.get_function("SpeakerWordsUDAF", database="multimodal_lake")
   
       from fabric_data.multimodal.function import AggregateFnBuilder
       udaf_builder = AggregateFnBuilder(
           fn=udaf,
           on=[ds.chapter_id, ds.hyp_text],
           as_col="tf_idf_image",
           num_dpus=0.5,
       )
   
       summary = ds.aggregate(
           udaf_builder,
           by=ds.chapter_id,
       )
       summary.show(limit=1)
   
       df = summary.execute()
       print(df)
       display_image(df["tf_idf_image"].tolist())