From 95bed2337e8065d3331109d6c2d00349ad82fd77 Mon Sep 17 00:00:00 2001
From: 游雁 <zhifu.gzf@alibaba-inc.com>
Date: 星期三, 08 五月 2024 19:14:21 +0800
Subject: [PATCH] Merge branch 'dev_gzf_exp' of github.com:alibaba-damo-academy/FunASR into dev_gzf_exp merge

---
 funasr/models/sense_voice/model.py  |   28 ++++++++++++++
 funasr/models/sense_voice/search.py |   59 +++++++++++++++++++++++++++++
 README_zh.md                        |    4 +-
 README.md                           |    4 +-
 4 files changed, 91 insertions(+), 4 deletions(-)

diff --git a/README.md b/README.md
index faa758c..8b093bc 100644
--- a/README.md
+++ b/README.md
@@ -83,8 +83,8 @@
 |                                   fsmn-vad <br> ( [猸怾(https://modelscope.cn/models/damo/speech_fsmn_vad_zh-cn-16k-common-pytorch/summary) [馃](https://huggingface.co/funasr/fsmn-vad) )                                   |               voice activity detection                | 5000 hours, Mandarin and English |    0.4M    | 
 |                                     fa-zh <br> ( [猸怾(https://modelscope.cn/models/damo/speech_timestamp_prediction-v1-16k-offline/summary) [馃](https://huggingface.co/funasr/fa-zh) )                                     |                 timestamp prediction                  |       5000 hours, Mandarin       |    38M     | 
 |                                       cam++ <br> ( [猸怾(https://modelscope.cn/models/iic/speech_campplus_sv_zh-cn_16k-common/summary) [馃](https://huggingface.co/funasr/campplus) )                                        |           speaker verification/diarization            |            5000 hours            |    7.2M    | 
-|                                                  Whisper-large-v2 <br> ([猸怾(https://www.modelscope.cn/models/iic/speech_whisper-large_asr_multilingual/summary)  [馃崁](https://github.com/openai/whisper) )                                                  |  speech recognition, with timestamps, non-streaming   |          multilingual            |    1.5G    |
-|                                                Whisper-large-v3 <br> ([猸怾(https://www.modelscope.cn/models/iic/Whisper-large-v3/summary)  [馃崁](https://github.com/openai/whisper) )                                                 |  speech recognition, with timestamps, non-streaming   |          multilingual            |    1.5G    |
+|                                                  Whisper-large-v2 <br> ([猸怾(https://www.modelscope.cn/models/iic/speech_whisper-large_asr_multilingual/summary)  [馃崁](https://github.com/openai/whisper) )                                                  |  speech recognition, with timestamps, non-streaming   |          multilingual      |    1550 M    |
+|                                                Whisper-large-v3 <br> ([猸怾(https://www.modelscope.cn/models/iic/Whisper-large-v3/summary)  [馃崁](https://github.com/openai/whisper) )                                                 |  speech recognition, with timestamps, non-streaming   |          multilingual            |    1550 M    |
 |                                         Qwen-Audio <br> ([猸怾(examples/industrial_data_pretraining/qwen_audio/demo.py)  [馃](https://huggingface.co/Qwen/Qwen-Audio) )                                         |      audio-text multimodal models (pretraining)       |     multilingual      |  8B  |
 |                   Qwen-Audio-Chat <br> ([猸怾(examples/industrial_data_pretraining/qwen_audio/demo_chat.py)  [馃](https://huggingface.co/Qwen/Qwen-Audio-Chat) )                                                |          audio-text multimodal models (chat)          |     multilingual      |  8B  |
 
diff --git a/README_zh.md b/README_zh.md
index 80c2e7e..963469a 100644
--- a/README_zh.md
+++ b/README_zh.md
@@ -80,11 +80,11 @@
 | paraformer-zh-streaming <br> ( [猸怾(https://modelscope.cn/models/damo/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-online/summary) [馃](https://huggingface.co/funasr/paraformer-zh-streaming) ) |      璇煶璇嗗埆锛屽疄鏃�       |  60000灏忔椂锛屼腑鏂�  | 220M |
 |         paraformer-en <br> ( [猸怾(https://www.modelscope.cn/models/damo/speech_paraformer-large-vad-punc_asr_nat-en-16k-common-vocab10020/summary) [馃](https://huggingface.co/funasr/paraformer-en) )         |      璇煶璇嗗埆锛岄潪瀹炴椂      |  50000灏忔椂锛岃嫳鏂�  | 220M |
 |                      conformer-en <br> ( [猸怾(https://modelscope.cn/models/damo/speech_conformer_asr-en-16k-vocab4199-pytorch/summary) [馃](https://huggingface.co/funasr/conformer-en) )                      |      璇煶璇嗗埆锛岄潪瀹炴椂      |  50000灏忔椂锛岃嫳鏂�  | 220M |
-|                        ct-punc <br> ( [猸怾(https://modelscope.cn/models/damo/punc_ct-transformer_cn-en-common-vocab471067-large/summary) [馃](https://huggingface.co/funasr/ct-punc) )                         |        鏍囩偣鎭㈠        |  100M锛屼腑鏂囦笌鑻辨枃  | 1.1G | 
+|                        ct-punc <br> ( [猸怾(https://modelscope.cn/models/damo/punc_ct-transformer_cn-en-common-vocab471067-large/summary) [馃](https://huggingface.co/funasr/ct-punc) )                         |        鏍囩偣鎭㈠        |  100M锛屼腑鏂囦笌鑻辨枃  | 1.1B | 
 |                            fsmn-vad <br> ( [猸怾(https://modelscope.cn/models/damo/speech_fsmn_vad_zh-cn-16k-common-pytorch/summary) [馃](https://huggingface.co/funasr/fsmn-vad) )                             |     璇煶绔偣妫�娴嬶紝瀹炴椂      | 5000灏忔椂锛屼腑鏂囦笌鑻辨枃 | 0.4M | 
 |                              fa-zh <br> ( [猸怾(https://modelscope.cn/models/damo/speech_timestamp_prediction-v1-16k-offline/summary) [馃](https://huggingface.co/funasr/fa-zh) )                               |      瀛楃骇鍒椂闂存埑棰勬祴      |  50000灏忔椂锛屼腑鏂�  | 38M  |
 |                                 cam++ <br> ( [猸怾(https://modelscope.cn/models/iic/speech_campplus_sv_zh-cn_16k-common/summary) [馃](https://huggingface.co/funasr/campplus) )                                 |      璇磋瘽浜虹‘璁�/鍒嗗壊      |    5000灏忔椂    | 7.2M | 
-|                                     Whisper-large-v3 <br> ([猸怾(https://www.modelscope.cn/models/iic/Whisper-large-v3/summary)  [馃崁](https://github.com/openai/whisper) )                                      |  璇煶璇嗗埆锛屽甫鏃堕棿鎴宠緭鍑猴紝闈炲疄鏃�   |     澶氳瑷�      |  1G  |
+|                                     Whisper-large-v3 <br> ([猸怾(https://www.modelscope.cn/models/iic/Whisper-large-v3/summary)  [馃崁](https://github.com/openai/whisper) )                                      |  璇煶璇嗗埆锛屽甫鏃堕棿鎴宠緭鍑猴紝闈炲疄鏃�   |     澶氳瑷�      |  1550 M  |
 |                                         Qwen-Audio <br> ([猸怾(examples/industrial_data_pretraining/qwen_audio/demo.py)  [馃](https://huggingface.co/Qwen/Qwen-Audio) )                                         |  闊抽鏂囨湰澶氭ā鎬佸ぇ妯″瀷锛堥璁粌锛�   |     澶氳瑷�      |  8B  |
 |                   Qwen-Audio-Chat <br> ([猸怾(examples/industrial_data_pretraining/qwen_audio/demo_chat.py)  [馃](https://huggingface.co/Qwen/Qwen-Audio-Chat) )                                                | 闊抽鏂囨湰澶氭ā鎬佸ぇ妯″瀷锛坈hat鐗堟湰锛� |     澶氳瑷�      |  8B  |
 
diff --git a/funasr/models/sense_voice/model.py b/funasr/models/sense_voice/model.py
index 0230638..00bc85b 100644
--- a/funasr/models/sense_voice/model.py
+++ b/funasr/models/sense_voice/model.py
@@ -514,6 +514,20 @@
         self.beam_search.sos = sos_int
         self.beam_search.eos = eos_int[0]
 
+        # Paramterts for rich decoding
+        self.beam_search.emo_unk = tokenizer.encode(
+            DecodingOptions.get("emo_unk_token", "<|SPECIAL_TOKEN_1|>"), allowed_special="all")[0]
+        self.beam_search.emo_unk_score = 1
+        self.beam_search.emo_tokens = tokenizer.encode(
+            DecodingOptions.get("emo_target_tokens", "<|HAPPY|><|SAD|><|ANGRY|>"), allowed_special="all")
+        self.beam_search.emo_scores = DecodingOptions.get("emo_target_threshold", [0.1, 0.1, 0.1])
+
+        self.beam_search.event_bg_token = tokenizer.encode(
+            DecodingOptions.get("gain_tokens_bg", "<|Speech|><|BGM|><|Applause|><|Laughter|>"), allowed_special="all")
+        self.beam_search.event_ed_token = tokenizer.encode(
+            DecodingOptions.get("gain_tokens_ed", "<|/Speech|><|/BGM|><|/Applause|><|/Laughter|>"), allowed_special="all")
+        self.beam_search.event_score_ga = DecodingOptions.get("gain_tokens_score", [1, 1, 1, 1])
+
         encoder_out, encoder_out_lens = self.encode(
             speech[None, :, :].permute(0, 2, 1), speech_lengths
         )
@@ -843,6 +857,20 @@
         self.beam_search.sos = sos_int
         self.beam_search.eos = eos_int[0]
 
+        # Paramterts for rich decoding
+        self.beam_search.emo_unk = tokenizer.encode(
+            DecodingOptions.get("emo_unk_token", "<|SPECIAL_TOKEN_1|>"), allowed_special="all")[0]
+        self.beam_search.emo_unk_score = 1
+        self.beam_search.emo_tokens = tokenizer.encode(
+            DecodingOptions.get("emo_target_tokens", "<|HAPPY|><|SAD|><|ANGRY|>"), allowed_special="all")
+        self.beam_search.emo_scores = DecodingOptions.get("emo_target_threshold", [0.1, 0.1, 0.1])
+
+        self.beam_search.event_bg_token = tokenizer.encode(
+            DecodingOptions.get("gain_tokens_bg", "<|Speech|><|BGM|><|Applause|><|Laughter|>"), allowed_special="all")
+        self.beam_search.event_ed_token = tokenizer.encode(
+            DecodingOptions.get("gain_tokens_ed", "<|/Speech|><|/BGM|><|/Applause|><|/Laughter|>"), allowed_special="all")
+        self.beam_search.event_score_ga = DecodingOptions.get("gain_tokens_score", [1, 1, 1, 1])
+
         encoder_out, encoder_out_lens = self.encode(
             speech[None, :, :].permute(0, 2, 1), speech_lengths
         )
diff --git a/funasr/models/sense_voice/search.py b/funasr/models/sense_voice/search.py
index 694e569..4400ce7 100644
--- a/funasr/models/sense_voice/search.py
+++ b/funasr/models/sense_voice/search.py
@@ -1,4 +1,5 @@
 from itertools import chain
+from dataclasses import field
 import logging
 from typing import Any
 from typing import Dict
@@ -8,6 +9,7 @@
 from typing import Union
 
 import torch
+import numpy as np
 
 from funasr.metrics.common import end_detect
 from funasr.models.transformer.scorers.scorer_interface import PartialScorerInterface
@@ -42,6 +44,17 @@
         vocab_size: int,
         sos=None,
         eos=None,
+        # NOTE add rich decoding parameters
+        # [SPECIAL_TOKEN_1, HAPPY, SAD, ANGRY, NEUTRAL]
+        emo_unk: int = 58964,
+        emo_unk_score: float = 1.0,
+        emo_tokens: List[int] = field(default_factory=lambda: [58954, 58955, 58956, 58957]),
+        emo_scores: List[float] = field(default_factory=lambda: [0.1, 0.1, 0.1, 0.1]),
+        # [Speech, BGM, Laughter, Applause]
+        event_bg_token: List[int] = field(default_factory=lambda: [58946, 58948, 58950, 58952]),
+        event_ed_token: List[int] = field(default_factory=lambda: [58947, 58949, 58951, 58953]),
+        event_score_ga: List[float] = field(default_factory=lambda: [1, 1, 5, 25]),
+
         token_list: List[str] = None,
         pre_beam_ratio: float = 1.5,
         pre_beam_score_key: str = None,
@@ -110,6 +123,14 @@
             and len(self.part_scorers) > 0
         )
 
+        self.emo_unk = emo_unk
+        self.emo_unk_score = emo_unk_score
+        self.emo_tokens = emo_tokens
+        self.emo_scores = emo_scores
+        self.event_bg_token = event_bg_token
+        self.event_ed_token = event_ed_token
+        self.event_score_ga = event_score_ga
+
     def init_hyp(self, x: torch.Tensor) -> List[Hypothesis]:
         """Get an initial hypothesis data.
 
@@ -170,10 +191,48 @@
         """
         scores = dict()
         states = dict()
+
+        def get_score(yseq, sp1, sp2):
+            score = [0, 0]
+            last_token = yseq[-1]
+            last_token2 = yseq[-2] if len(yseq) > 1 else yseq[-1]
+            sum_sp1 = sum([1 if x == sp1 else 0 for x in yseq])
+            sum_sp2 = sum([1 if x == sp2 else 0 for x in yseq])
+            if sum_sp1 > sum_sp2 or last_token in [sp1, sp2]:
+                score[0] = -np.inf
+            if sum_sp2 >= sum_sp1:
+                score[1] = -np.inf
+            return score
+
+        def struct_score(yseq, score):
+            import math
+
+            last_token = yseq[-1]
+            if last_token in self.emo_tokens + [self.emo_unk]:
+                # prevent output event after emotation token 
+                score[self.event_bg_token] = -np.inf
+
+            for eve_bg, eve_ed, eve_ga in zip(self.event_bg_token, self.event_ed_token, self.event_score_ga):
+                score_offset = get_score(yseq, eve_bg, eve_ed)
+                score[eve_bg] += score_offset[0]
+                score[eve_ed] += score_offset[1]
+                score[eve_bg] += math.log(eve_ga)
+
+
+            score[self.emo_unk] += math.log(self.emo_unk_score)
+            for emo, emo_th in zip(self.emo_tokens, self.emo_scores):
+                if score.argmax() == emo and score[emo] < math.log(emo_th):
+                    score[self.emo_unk] = max(score[emo], score[self.emo_unk])
+                    score[emo] = -np.inf
+            return score
+
         for k, d in self.full_scorers.items():
             scores[k], states[k] = d.score(hyp.yseq, hyp.states[k], x)
+            scores[k] = struct_score(hyp.yseq, scores[k])
+
         return scores, states
 
+
     def score_partial(
         self, hyp: Hypothesis, ids: torch.Tensor, x: torch.Tensor
     ) -> Tuple[Dict[str, torch.Tensor], Dict[str, Any]]:

--
Gitblit v1.9.1