walston commited on
Commit
bf495c7
·
verified ·
1 Parent(s): 2240d09

Inference via the xmdd package; drop unused tokens.txt

Browse files
Files changed (3) hide show
  1. README.md +14 -12
  2. config.yaml +0 -1
  3. tokens.txt +0 -188
README.md CHANGED
@@ -38,31 +38,33 @@ the learner actually produced. Aligning it to the canonical phones gives detecti
38
 
39
  | File | Content |
40
  |---|---|
41
- | `model.safetensors` | Weights (bf16; keys `asr.*` = Qwen3-ASR, `ctc_head.*` = inherited unused CTC head), optimizer state removed |
42
  | `config.yaml` | Inference config (prompt layout, language prefixes, model settings) |
43
  | `processor/` | Qwen3-ASR processor and tokenizer, including the added prompt boundary tokens |
44
  | `phone_vocabulary.txt` | The 242-phone training vocabulary (zh tonal phones + CMU 39 phones), in index order |
45
- | `tokens.txt` | Base phone token table |
46
  | `arabic_phone_symbols.txt` | Phone symbols used for constrained decoding on Arabic (QuranMB) |
47
 
48
  ## Usage
49
 
50
- Inference uses the code in the X-MDD repository (`scripts/evaluate_qwen3_multilingual_bpe_icl_dev.py`),
51
- which reads this folder directly. The base `Qwen/Qwen3-ASR-1.7B-hf` is downloaded for the architecture
52
- and then overwritten by these weights.
53
 
54
  ```bash
55
  hf download walston/X-MDD --local-dir pretrained/X-MDD
56
- M=pretrained/X-MDD
57
- python scripts/evaluate_qwen3_multilingual_bpe_icl_dev.py \
58
- --config $M/config.yaml --checkpoint $M/model.safetensors \
59
- --languages en --split test --manifest my_en_test.jsonl \
60
- --shots 4 --demo-source test --skip-loss \
61
  --output results/en.json --predictions results/en.pred.jsonl
62
  ```
63
 
64
- For Arabic (or another unseen language) add `--raw-phones --constrain-phones <manifest>`; see the
65
- repository README for all options.
 
 
 
 
 
 
 
 
66
 
67
  Input manifest rows (JSONL) look like:
68
 
 
38
 
39
  | File | Content |
40
  |---|---|
41
+ | `model.safetensors` | Weights (bf16; keys `asr.*` = Qwen3-ASR; `ctc_head.*` = an unused CTC head, ignored at inference) |
42
  | `config.yaml` | Inference config (prompt layout, language prefixes, model settings) |
43
  | `processor/` | Qwen3-ASR processor and tokenizer, including the added prompt boundary tokens |
44
  | `phone_vocabulary.txt` | The 242-phone training vocabulary (zh tonal phones + CMU 39 phones), in index order |
 
45
  | `arabic_phone_symbols.txt` | Phone symbols used for constrained decoding on Arabic (QuranMB) |
46
 
47
  ## Usage
48
 
49
+ Install the [X-MDD code](https://github.com/oshindow/whisper_pinyin) (`xmdd` package), then:
 
 
50
 
51
  ```bash
52
  hf download walston/X-MDD --local-dir pretrained/X-MDD
53
+ python -m xmdd.infer --model-dir pretrained/X-MDD --language en \
54
+ --manifest my_en_test.jsonl --shots 4 \
 
 
 
55
  --output results/en.json --predictions results/en.pred.jsonl
56
  ```
57
 
58
+ ```python
59
+ from xmdd import XMDD, load_manifest, select_demos
60
+ model = XMDD.from_pretrained("pretrained/X-MDD")
61
+ queries = load_manifest("my_en_test.jsonl", "en", model.cfg)
62
+ print(model.predict(queries[0], select_demos(queries[0], queries, shots=4)))
63
+ ```
64
+
65
+ For Arabic or another unseen language add `--raw-phones --constrain-phones <phones.txt>` (and
66
+ `--language-name` for a language other than zh/en/ar). The base `Qwen/Qwen3-ASR-1.7B-hf` is downloaded for
67
+ the architecture and then overwritten by these weights.
68
 
69
  Input manifest rows (JSONL) look like:
70
 
config.yaml CHANGED
@@ -25,7 +25,6 @@ data:
25
  audio_field: l2_wav
26
  actual_phone_field: actual_phones
27
  canonical_phone_field: canonical_phones
28
- token_table_path: tokens.txt
29
  sample_rate: 16000
30
  remove_phones:
31
  - sil
 
25
  audio_field: l2_wav
26
  actual_phone_field: actual_phones
27
  canonical_phone_field: canonical_phones
 
28
  sample_rate: 16000
29
  remove_phones:
30
  - sil
tokens.txt DELETED
@@ -1,188 +0,0 @@
1
- <blk> 0
2
- <sos/eos> 1
3
- <unk> 2
4
- a1 3
5
- a2 4
6
- a3 5
7
- a4 6
8
- a5 7
9
- aa 8
10
- ai1 9
11
- ai2 10
12
- ai3 11
13
- ai4 12
14
- ai5 13
15
- an1 14
16
- an2 15
17
- an3 16
18
- an4 17
19
- ang1 18
20
- ang2 19
21
- ang3 20
22
- ang4 21
23
- ao1 22
24
- ao2 23
25
- ao3 24
26
- ao4 25
27
- b 26
28
- c 27
29
- ch 28
30
- d 29
31
- e1 30
32
- e2 31
33
- e3 32
34
- e4 33
35
- e5 34
36
- ee 35
37
- ei1 36
38
- ei2 37
39
- ei3 38
40
- ei4 39
41
- ei5 40
42
- en1 41
43
- en2 42
44
- en3 43
45
- en4 44
46
- en5 45
47
- eng1 46
48
- eng2 47
49
- eng3 48
50
- eng4 49
51
- er2 50
52
- er3 51
53
- er4 52
54
- er5 53
55
- f 54
56
- g 55
57
- h 56
58
- i1 57
59
- i2 58
60
- i3 59
61
- i4 60
62
- i5 61
63
- ia1 62
64
- ia2 63
65
- ia3 64
66
- ia4 65
67
- ia5 66
68
- ian1 67
69
- ian2 68
70
- ian3 69
71
- ian4 70
72
- ian5 71
73
- iang1 72
74
- iang2 73
75
- iang3 74
76
- iang4 75
77
- iang5 76
78
- iao1 77
79
- iao2 78
80
- iao3 79
81
- iao4 80
82
- ie1 81
83
- ie2 82
84
- ie3 83
85
- ie4 84
86
- ie5 85
87
- ii 86
88
- in1 87
89
- in2 88
90
- in3 89
91
- in4 90
92
- ing1 91
93
- ing2 92
94
- ing3 93
95
- ing4 94
96
- iong1 95
97
- iong2 96
98
- iong3 97
99
- iong4 98
100
- iu1 99
101
- iu2 100
102
- iu3 101
103
- iu4 102
104
- iu5 103
105
- iz4 104
106
- j 105
107
- k 106
108
- l 107
109
- m 108
110
- n 109
111
- o1 110
112
- o2 111
113
- o3 112
114
- o4 113
115
- o5 114
116
- ong1 115
117
- ong2 116
118
- ong3 117
119
- ong4 118
120
- oo 119
121
- ou1 120
122
- ou2 121
123
- ou3 122
124
- ou4 123
125
- ou5 124
126
- p 125
127
- q 126
128
- r 127
129
- s 128
130
- sh 129
131
- t 130
132
- u1 131
133
- u2 132
134
- u3 133
135
- u4 134
136
- u5 135
137
- ua1 136
138
- ua2 137
139
- ua3 138
140
- ua4 139
141
- ua5 140
142
- uai1 141
143
- uai2 142
144
- uai3 143
145
- uai4 144
146
- uan1 145
147
- uan2 146
148
- uan3 147
149
- uan4 148
150
- uang1 149
151
- uang2 150
152
- uang3 151
153
- uang4 152
154
- ueng1 153
155
- ui1 154
156
- ui2 155
157
- ui3 156
158
- ui4 157
159
- un1 158
160
- un2 159
161
- un3 160
162
- un4 161
163
- uo1 162
164
- uo2 163
165
- uo3 164
166
- uo4 165
167
- uo5 166
168
- uu 167
169
- v1 168
170
- v2 169
171
- v3 170
172
- v4 171
173
- van1 172
174
- van2 173
175
- van3 174
176
- van4 175
177
- ve1 176
178
- ve2 177
179
- ve3 178
180
- ve4 179
181
- vn1 180
182
- vn2 181
183
- vn3 182
184
- vn4 183
185
- vv 184
186
- x 185
187
- z 186
188
- zh 187