Paul Bird
commited on
Upload 5 files
Browse files- .gitattributes +3 -0
- AudioDecoder_Tiny.sentis +3 -0
- AudioEncoder_Tiny.sentis +3 -0
- LogMelSepctro.sentis +3 -0
- RunWhisper.cs +213 -0
- vocab.json +0 -0
.gitattributes
CHANGED
@@ -33,3 +33,6 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
|
|
33 |
*.zip filter=lfs diff=lfs merge=lfs -text
|
34 |
*.zst filter=lfs diff=lfs merge=lfs -text
|
35 |
*tfevents* filter=lfs diff=lfs merge=lfs -text
|
|
|
|
|
|
|
|
33 |
*.zip filter=lfs diff=lfs merge=lfs -text
|
34 |
*.zst filter=lfs diff=lfs merge=lfs -text
|
35 |
*tfevents* filter=lfs diff=lfs merge=lfs -text
|
36 |
+
AudioDecoder_Tiny.sentis filter=lfs diff=lfs merge=lfs -text
|
37 |
+
AudioEncoder_Tiny.sentis filter=lfs diff=lfs merge=lfs -text
|
38 |
+
LogMelSepctro.sentis filter=lfs diff=lfs merge=lfs -text
|
AudioDecoder_Tiny.sentis
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:f6d24553eda46f335ead8ba30e3970fc8056086a538047248821aa31a135f938
|
3 |
+
size 198832845
|
AudioEncoder_Tiny.sentis
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:d3fb532b04b438079db8de9551a0d813da22be5fd05cdeeff3d09794492ca5b1
|
3 |
+
size 32888514
|
LogMelSepctro.sentis
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:e021007141fdf2d39113ea1aa12bc258226ea1c2976171544f3a05979e2b69ef
|
3 |
+
size 1360848
|
RunWhisper.cs
ADDED
@@ -0,0 +1,213 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
using System.Collections;
|
2 |
+
using System.Collections.Generic;
|
3 |
+
using UnityEngine;
|
4 |
+
using Unity.Sentis;
|
5 |
+
using System.IO;
|
6 |
+
using Newtonsoft.Json;
|
7 |
+
using System.Text;
|
8 |
+
|
9 |
+
/*
|
10 |
+
* Whisper Inference Code
|
11 |
+
* ======================
|
12 |
+
*
|
13 |
+
* Put this script on the Main Camera
|
14 |
+
*
|
15 |
+
* In Assets/StreamingAssets put:
|
16 |
+
*
|
17 |
+
* AudioDecoder_Tiny.sentis
|
18 |
+
* AudioEncoder_Tiny.sentis
|
19 |
+
* LogMelSepctro.sentis
|
20 |
+
* vocab.json
|
21 |
+
*
|
22 |
+
* Drag a 30s 16khz mono uncompressed audioclip into the audioClip field.
|
23 |
+
*
|
24 |
+
* Install package com.unity.nuget.newtonsoft-json from packagemanger
|
25 |
+
* Install package com.unity.sentis
|
26 |
+
*
|
27 |
+
*/
|
28 |
+
|
29 |
+
|
30 |
+
public class RunWhisper : MonoBehaviour
|
31 |
+
{
|
32 |
+
IWorker decoderEngine, encoderEngine, spectroEngine;
|
33 |
+
|
34 |
+
const BackendType backend = BackendType.GPUCompute;
|
35 |
+
|
36 |
+
// Link your audioclip here. Format must be 16Hz mono non-compressed.
|
37 |
+
public AudioClip audioClip;
|
38 |
+
|
39 |
+
const int maxTokens = 100;
|
40 |
+
|
41 |
+
//Special tokens
|
42 |
+
const int END_OF_TEXT = 50257;
|
43 |
+
const int START_OF_TRANSCRIPT = 50258;
|
44 |
+
const int ENGLISH = 50259;
|
45 |
+
const int TRANSCRIBE = 50359;
|
46 |
+
const int START_TIME = 50364;
|
47 |
+
|
48 |
+
Ops ops;
|
49 |
+
ITensorAllocator allocator;
|
50 |
+
|
51 |
+
int numSamples;
|
52 |
+
float[] data;
|
53 |
+
string[] tokens;
|
54 |
+
|
55 |
+
int currentToken = 0;
|
56 |
+
int[] outputTokens = new int[maxTokens];
|
57 |
+
|
58 |
+
// Used for special character decoding
|
59 |
+
int[] shiftDownDict = new int[256];
|
60 |
+
|
61 |
+
TensorFloat encodedAudio;
|
62 |
+
|
63 |
+
bool transcribe = false;
|
64 |
+
string outputString = "";
|
65 |
+
|
66 |
+
void Start()
|
67 |
+
{
|
68 |
+
allocator = new TensorCachingAllocator();
|
69 |
+
ops = WorkerFactory.CreateOps(backend, allocator);
|
70 |
+
|
71 |
+
SetupCharacterShifts();
|
72 |
+
|
73 |
+
GetTokens();
|
74 |
+
|
75 |
+
Model decoder = ModelLoader.Load(Application.streamingAssetsPath + "/AudioDecoder_Tiny.sentis");
|
76 |
+
Model encoder = ModelLoader.Load(Application.streamingAssetsPath + "/AudioEncoder_Tiny.sentis");
|
77 |
+
Model spectro = ModelLoader.Load(Application.streamingAssetsPath + "/LogMelSepctro.sentis");
|
78 |
+
|
79 |
+
decoderEngine = WorkerFactory.CreateWorker(backend, decoder);
|
80 |
+
encoderEngine = WorkerFactory.CreateWorker(backend, encoder);
|
81 |
+
spectroEngine = WorkerFactory.CreateWorker(backend, spectro);
|
82 |
+
|
83 |
+
outputTokens[0] = START_OF_TRANSCRIPT;
|
84 |
+
outputTokens[1] = ENGLISH;
|
85 |
+
outputTokens[2] = TRANSCRIBE;
|
86 |
+
outputTokens[3] = START_TIME;
|
87 |
+
currentToken = 3;
|
88 |
+
|
89 |
+
LoadAudio();
|
90 |
+
EncodeAudio();
|
91 |
+
transcribe = true;
|
92 |
+
}
|
93 |
+
|
94 |
+
void LoadAudio()
|
95 |
+
{
|
96 |
+
if(audioClip.frequency != 16000)
|
97 |
+
{
|
98 |
+
Debug.Log($"The audio clip should have frequency 16kHz. It has frequency {audioClip.frequency / 1000f}kHz");
|
99 |
+
}
|
100 |
+
|
101 |
+
numSamples = audioClip.samples;
|
102 |
+
data = new float[numSamples];
|
103 |
+
audioClip.GetData(data, 0);
|
104 |
+
}
|
105 |
+
|
106 |
+
|
107 |
+
void GetTokens()
|
108 |
+
{
|
109 |
+
var jsonText = File.ReadAllText(Application.streamingAssetsPath + "/vocab.json");
|
110 |
+
var vocab = Newtonsoft.Json.JsonConvert.DeserializeObject<Dictionary<string, int>>(jsonText);
|
111 |
+
tokens = new string[vocab.Count];
|
112 |
+
foreach(var item in vocab)
|
113 |
+
{
|
114 |
+
tokens[item.Value] = item.Key;
|
115 |
+
}
|
116 |
+
}
|
117 |
+
|
118 |
+
void EncodeAudio()
|
119 |
+
{
|
120 |
+
var input = new TensorFloat(new TensorShape(1, numSamples), data);
|
121 |
+
|
122 |
+
int maxSamples = 30 * 16000;
|
123 |
+
if (numSamples > maxSamples)
|
124 |
+
{
|
125 |
+
Debug.Log("The AudioClip is too long.");
|
126 |
+
return;
|
127 |
+
}
|
128 |
+
|
129 |
+
// Pad out to 30 seconds at 16khz if necessary
|
130 |
+
var input30seconds = ops.Pad(input, new int[] { 0, 0, 0, 30 * 16000 - numSamples });
|
131 |
+
|
132 |
+
spectroEngine.Execute(input30seconds);
|
133 |
+
var spectroOutput = spectroEngine.PeekOutput() as TensorFloat;
|
134 |
+
|
135 |
+
encoderEngine.Execute(spectroOutput);
|
136 |
+
encodedAudio = encoderEngine.PeekOutput() as TensorFloat;
|
137 |
+
}
|
138 |
+
|
139 |
+
|
140 |
+
// Update is called once per frame
|
141 |
+
void Update()
|
142 |
+
{
|
143 |
+
if (transcribe && currentToken < outputTokens.Length - 1)
|
144 |
+
{
|
145 |
+
var tokensSoFar = new TensorInt(new TensorShape(1, outputTokens.Length), outputTokens);
|
146 |
+
|
147 |
+
var inputs = new Dictionary<string, Tensor>
|
148 |
+
{
|
149 |
+
{"encoded_audio",encodedAudio },
|
150 |
+
{"tokens" , tokensSoFar }
|
151 |
+
};
|
152 |
+
|
153 |
+
decoderEngine.Execute(inputs);
|
154 |
+
var tokensOut = decoderEngine.PeekOutput() as TensorFloat;
|
155 |
+
|
156 |
+
var tokensPredictions = ops.ArgMax(tokensOut, 2, false);
|
157 |
+
tokensPredictions.MakeReadable();
|
158 |
+
|
159 |
+
int ID = tokensPredictions[currentToken];
|
160 |
+
|
161 |
+
currentToken++;
|
162 |
+
outputTokens[currentToken] = ID;
|
163 |
+
|
164 |
+
if (ID == END_OF_TEXT)
|
165 |
+
{
|
166 |
+
transcribe = false;
|
167 |
+
}
|
168 |
+
else if (ID >= tokens.Length) outputString += $"(time={(ID - START_TIME) * 0.02f})";
|
169 |
+
else outputString += GetUnicodeText(tokens[ID]);
|
170 |
+
|
171 |
+
Debug.Log(outputString);
|
172 |
+
}
|
173 |
+
}
|
174 |
+
|
175 |
+
// Translates encoded special characters to Unicode
|
176 |
+
string GetUnicodeText(string text)
|
177 |
+
{
|
178 |
+
var bytes = Encoding.GetEncoding("ISO-8859-1").GetBytes(ShiftCharacterDown(text));
|
179 |
+
return Encoding.UTF8.GetString(bytes);
|
180 |
+
}
|
181 |
+
|
182 |
+
string ShiftCharacterDown(string text)
|
183 |
+
{
|
184 |
+
string outText = "";
|
185 |
+
foreach (char letter in text)
|
186 |
+
{
|
187 |
+
outText += ((int)letter <= 256) ? letter :
|
188 |
+
(char)shiftDownDict[(int)(letter - 256)];
|
189 |
+
}
|
190 |
+
return outText;
|
191 |
+
}
|
192 |
+
|
193 |
+
void SetupCharacterShifts()
|
194 |
+
{
|
195 |
+
for (int i = 0, n = 0; i < 256; i++)
|
196 |
+
{
|
197 |
+
if (IsWhiteSpace((char)i)) shiftDownDict[n++] = i;
|
198 |
+
}
|
199 |
+
}
|
200 |
+
|
201 |
+
bool IsWhiteSpace(char c)
|
202 |
+
{
|
203 |
+
return !(('!' <= c && c <= '~') || ('�' <= c && c <= '�') || ('�' <= c && c <= '�'));
|
204 |
+
}
|
205 |
+
|
206 |
+
private void OnDestroy()
|
207 |
+
{
|
208 |
+
decoderEngine?.Dispose();
|
209 |
+
encoderEngine?.Dispose();
|
210 |
+
spectroEngine?.Dispose();
|
211 |
+
ops?.Dispose();
|
212 |
+
}
|
213 |
+
}
|
vocab.json
ADDED
The diff for this file is too large to render.
See raw diff
|
|