Hugging Face's logo

Spaces:
Luigi
/
Streaming-Zipformer
Running

App Files Community
Streaming-Zipformer
File size: 8,580 Bytes 
4a7594f
454a10d
2318eae
 
221a9c5
231cd3a
ab9b679
4a7594f
 
 
 
231cd3a
 
 
454a10d
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2632b6f
454a10d
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
ab9b679
 
b26d94c
 
 
ab9b679
454a10d
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2318eae
454a10d
 
 
 
2318eae
 
 
 
 
 
 
7c3f2af
2318eae
2a8a9a5
 
 
7c3f2af
2a8a9a5
 
 
2318eae
 
 
231cd3a
2318eae
 
 
 
 
 
 
231cd3a
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
 
import os
from pathlib import Path
import numpy as np
import sherpa_onnx
import scipy.signal
from opencc import OpenCC
from huggingface_hub import hf_hub_download

# Ensure Hugging Face cache is in a user-writable directory
CACHE_DIR = Path(__file__).parent / "hf_cache"
os.makedirs(CACHE_DIR, exist_ok=True)

converter = OpenCC('s2t')

# Streaming Zipformer model registry: paths relative to repo root
STREAMING_ZIPFORMER_MODELS = {
    "pfluo/k2fsa-zipformer-chinese-english-mixed": {
        "tokens": "data/lang_char_bpe/tokens.txt",
        "encoder_fp32": "exp/encoder-epoch-99-avg-1.onnx",
        "encoder_int8": "exp/encoder-epoch-99-avg-1.int8.onnx",
        "decoder_fp32": "exp/decoder-epoch-99-avg-1.onnx",
        "decoder_int8": None,
        "joiner_fp32": "exp/joiner-epoch-99-avg-1.onnx",
        "joiner_int8": "exp/joiner-epoch-99-avg-1.int8.onnx",
    },
    "k2-fsa/sherpa-onnx-streaming-zipformer-korean-2024-06-16": {
        "tokens": "tokens.txt",
        "encoder_fp32": "encoder-epoch-99-avg-1.onnx",
        "encoder_int8": "encoder-epoch-99-avg-1.int8.onnx",
        "decoder_fp32": "decoder-epoch-99-avg-1.onnx",
        "decoder_int8": "decoder-epoch-99-avg-1.int8.onnx",
        "joiner_fp32": "joiner-epoch-99-avg-1.onnx",
        "joiner_int8": "joiner-epoch-99-avg-1.int8.onnx",
    },
    "k2-fsa/sherpa-onnx-streaming-zipformer-multi-zh-hans-2023-12-12": {
        "tokens": "tokens.txt",
        "encoder_fp32": "encoder-epoch-20-avg-1-chunk-16-left-128.onnx",
        "encoder_int8": "encoder-epoch-20-avg-1-chunk-16-left-128.int8.onnx",
        "decoder_fp32": "decoder-epoch-20-avg-1-chunk-16-left-128.onnx",
        "decoder_int8": "decoder-epoch-20-avg-1-chunk-16-left-128.int8.onnx",
        "joiner_fp32": "joiner-epoch-20-avg-1-chunk-16-left-128.onnx",
        "joiner_int8": "joiner-epoch-20-avg-1-chunk-16-left-128.int8.onnx",
    },
    "pkufool/icefall-asr-zipformer-streaming-wenetspeech-20230615": {
        "tokens": "data/lang_char/tokens.txt",
        "encoder_fp32": "exp/encoder-epoch-12-avg-4-chunk-16-left-128.onnx",
        "encoder_int8": "exp/encoder-epoch-12-avg-4-chunk-16-left-128.int8.onnx",
        "decoder_fp32": "exp/decoder-epoch-12-avg-4-chunk-16-left-128.onnx",
        "decoder_int8": "exp/decoder-epoch-12-avg-4-chunk-16-left-128.int8.onnx",
        "joiner_fp32": "exp/joiner-epoch-12-avg-4-chunk-16-left-128.onnx",
        "joiner_int8": "exp/joiner-epoch-12-avg-4-chunk-16-left-128.int8.onnx",
    },
    "csukuangfj/sherpa-onnx-streaming-zipformer-en-2023-06-26": {
        "tokens": "tokens.txt",
        "encoder_fp32": "encoder-epoch-99-avg-1-chunk-16-left-128.onnx",
        "encoder_int8": "encoder-epoch-99-avg-1-chunk-16-left-128.int8.onnx",
        "decoder_fp32": "decoder-epoch-99-avg-1-chunk-16-left-128.onnx",
        "decoder_int8": None,
        "joiner_fp32": "joiner-epoch-99-avg-1-chunk-16-left-128.onnx",
        "joiner_int8": "joiner-epoch-99-avg-1-chunk-16-left-128.int8.onnx",
    },
    "csukuangfj/sherpa-onnx-streaming-zipformer-en-2023-06-21": {
        "tokens": "tokens.txt",
        "encoder_fp32": "encoder-epoch-99-avg-1.onnx",
        "encoder_int8": "encoder-epoch-99-avg-1.int8.onnx",
        "decoder_fp32": "decoder-epoch-99-avg-1.onnx",
        "decoder_int8": "decoder-epoch-99-avg-1.int8.onnx",
        "joiner_fp32": "joiner-epoch-99-avg-1.onnx",
        "joiner_int8": "joiner-epoch-99-avg-1.int8.onnx",
    },
    "csukuangfj/sherpa-onnx-streaming-zipformer-en-2023-02-21": {
        "tokens": "tokens.txt",
        "encoder_fp32": "encoder-epoch-99-avg-1.onnx",
        "encoder_int8": "encoder-epoch-99-avg-1.int8.onnx",
        "decoder_fp32": "decoder-epoch-99-avg-1.onnx",
        "decoder_int8": "decoder-epoch-99-avg-1.int8.onnx",
        "joiner_fp32": "joiner-epoch-99-avg-1.onnx",
        "joiner_int8": "joiner-epoch-99-avg-1.int8.onnx",
    },
    "csukuangfj/sherpa-onnx-streaming-zipformer-bilingual-zh-en-2023-02-20": {
        "tokens": "tokens.txt",
        "encoder_fp32": "encoder-epoch-99-avg-1.onnx",
        "encoder_int8": "encoder-epoch-99-avg-1.int8.onnx",
        "decoder_fp32": "decoder-epoch-99-avg-1.onnx",
        "decoder_int8": "decoder-epoch-99-avg-1.int8.onnx",
        "joiner_fp32": "joiner-epoch-99-avg-1.onnx",
        "joiner_int8": "joiner-epoch-99-avg-1.int8.onnx",
    },
    "shaojieli/sherpa-onnx-streaming-zipformer-fr-2023-04-14": {
        "tokens": "tokens.txt",
        "encoder_fp32": "encoder-epoch-29-avg-9-with-averaged-model.onnx",
        "encoder_int8": "encoder-epoch-29-avg-9-with-averaged-model.int8.onnx",
        "decoder_fp32": "decoder-epoch-29-avg-9-with-averaged-model.onnx",
        "decoder_int8": "decoder-epoch-29-avg-9-with-averaged-model.int8.onnx",
        "joiner_fp32": "joiner-epoch-29-avg-9-with-averaged-model.onnx",
        "joiner_int8": "joiner-epoch-29-avg-9-with-averaged-model.int8.onnx",
    },
    "csukuangfj/sherpa-onnx-streaming-zipformer-small-bilingual-zh-en-2023-02-16": {
        "tokens": "tokens.txt",
        "encoder_fp32": "encoder-epoch-99-avg-1.onnx",
        "encoder_int8": "encoder-epoch-99-avg-1.int8.onnx",
        "decoder_fp32": "decoder-epoch-99-avg-1.onnx",
        "decoder_int8": "decoder-epoch-99-avg-1.int8.onnx",
        "joiner_fp32": "joiner-epoch-99-avg-1.onnx",
        "joiner_int8": "joiner-epoch-99-avg-1.int8.onnx",
    },
    "csukuangfj/sherpa-onnx-streaming-zipformer-zh-14M-2023-02-23": {
        "tokens": "tokens.txt",
        "encoder_fp32": "encoder-epoch-99-avg-1.onnx",
        "encoder_int8": "encoder-epoch-99-avg-1.int8.onnx",
        "decoder_fp32": "decoder-epoch-99-avg-1.onnx",
        "decoder_int8": "decoder-epoch-99-avg-1.int8.onnx",
        "joiner_fp32": "joiner-epoch-99-avg-1.onnx",
        "joiner_int8": "joiner-epoch-99-avg-1.int8.onnx",
    },
    "csukuangfj/sherpa-onnx-streaming-zipformer-en-20M-2023-02-17": {
        "tokens": "tokens.txt",
        "encoder_fp32": "encoder-epoch-99-avg-1.onnx",
        "encoder_int8": "encoder-epoch-99-avg-1.int8.onnx",
        "decoder_fp32": "decoder-epoch-99-avg-1.onnx",
        "decoder_int8": "decoder-epoch-99-avg-1.int8.onnx",
        "joiner_fp32": "joiner-epoch-99-avg-1.onnx",
        "joiner_int8": "joiner-epoch-99-avg-1.int8.onnx",
    },
}

# Audio resampling utility
def resample_audio(audio: np.ndarray, orig_sr: int, target_sr: int) -> np.ndarray:
    return scipy.signal.resample_poly(audio, target_sr, orig_sr)

# Create an online recognizer for a given model and precision
# model_id: full HF repo ID
# precision: "int8" or "fp32"
def create_recognizer(model_id: str, precision: str):
    if model_id not in STREAMING_ZIPFORMER_MODELS:
        raise ValueError(f"Model '{model_id}' is not registered.")
    entry = STREAMING_ZIPFORMER_MODELS[model_id]

    tokens_file = entry['tokens']
    encoder_file = entry['encoder_int8'] if precision == 'int8' else entry['encoder_fp32']
    decoder_file = entry['decoder_fp32']
    joiner_file = entry['joiner_int8'] if precision == 'int8' else entry['joiner_fp32']

    tokens_path = hf_hub_download(repo_id=model_id, filename=tokens_file, cache_dir=str(CACHE_DIR))
    encoder_path = hf_hub_download(repo_id=model_id, filename=encoder_file, cache_dir=str(CACHE_DIR))
    decoder_path = hf_hub_download(repo_id=model_id, filename=decoder_file, cache_dir=str(CACHE_DIR))
    joiner_path = hf_hub_download(repo_id=model_id, filename=joiner_file, cache_dir=str(CACHE_DIR))

    return sherpa_onnx.OnlineRecognizer.from_transducer(
        tokens=tokens_path,
        encoder=encoder_path,
        decoder=decoder_path,
        joiner=joiner_path,
        provider="cpu",
        num_threads=1,
        sample_rate=16000,
        feature_dim=80,
        decoding_method="greedy_search"
    )

def stream_audio(raw_pcm_bytes, stream, recognizer, orig_sr):
    audio = np.frombuffer(raw_pcm_bytes, dtype=np.float32)
    if audio.size == 0:
        return "", 0.0

    resampled = resample_audio(audio, orig_sr, 16000)
    rms = float(np.sqrt(np.mean(resampled ** 2)))

    stream.accept_waveform(16000, resampled)
    if recognizer.is_ready(stream):
        recognizer.decode_streams([stream])
    result = recognizer.get_result(stream)
    return converter.convert(result), rms

def finalize_stream(stream, recognizer):
    tail = np.zeros(int(0.66 * 16000), dtype=np.float32)
    stream.accept_waveform(16000, tail)
    stream.input_finished()
    while recognizer.is_ready(stream):
        recognizer.decode_streams([stream])
    result = recognizer.get_result(stream)
    return converter.convert(result)