@jskit-ai/assistant-voice-sherpa 0.1.1
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/README.md +171 -0
- package/bin/jskit-assistant-voice.js +7 -0
- package/models/streaming-zipformer-en-2023-06-26.bpe.vocab +500 -0
- package/models/voice-hotwords.txt +48 -0
- package/models/voice-models-kitten.json +116 -0
- package/models/voice-models-kokoro.json +102 -0
- package/models/voice-models-piper.json +200 -0
- package/models/voice-models.json +49 -0
- package/package.json +62 -0
- package/src/cli.js +167 -0
- package/src/index.js +2 -0
- package/src/sherpaSpeechEngine.js +292 -0
- package/src/synthesisProcess.js +81 -0
- package/src/synthesisWorker.js +40 -0
- package/src/voiceModelContract.js +117 -0
- package/tooling/build-voice-model-contract.mjs +51 -0
- package/tooling/prepare-models.sh +152 -0
- package/tooling/stage-voice-models.mjs +20 -0
- package/tooling/verify-voice-models.mjs +20 -0
package/README.md
ADDED
|
@@ -0,0 +1,171 @@
|
|
|
1
|
+
# Assistant voice with Sherpa
|
|
2
|
+
|
|
3
|
+
Optional local recognition and synthesis for `@jskit-ai/assistant-voice`.
|
|
4
|
+
The native npm dependency is `sherpa-onnx-node`; models are prepared explicitly.
|
|
5
|
+
No native dependency or model download is added to applications that only install
|
|
6
|
+
the browser/proxy package.
|
|
7
|
+
|
|
8
|
+
## Prepare and serve
|
|
9
|
+
|
|
10
|
+
Install `@jskit-ai/assistant-voice-sherpa`. Preparation requires Bash, tar with
|
|
11
|
+
bzip2, sha256sum, Node and curl (unless all pinned archives are already cached).
|
|
12
|
+
This extraction has been exercised on Linux x64 with Node 26. See the native
|
|
13
|
+
package's supported platforms before choosing another deployment target.
|
|
14
|
+
|
|
15
|
+
```sh
|
|
16
|
+
npx jskit-assistant-voice prepare --models-root /srv/speech/models --pack cori
|
|
17
|
+
npx jskit-assistant-voice verify --models-root /srv/speech/models
|
|
18
|
+
npx jskit-assistant-voice serve --models-root /srv/speech/models --key-file /run/secrets/speech-key
|
|
19
|
+
```
|
|
20
|
+
|
|
21
|
+
The key is a server-only random secret, at least 32 bytes. Create it once, with
|
|
22
|
+
restrictive permissions. Issue a separate tenant token for the application's
|
|
23
|
+
server and redirect it into its own protected file:
|
|
24
|
+
|
|
25
|
+
```sh
|
|
26
|
+
npx jskit-assistant-voice token --key-file /run/secrets/speech-key --tenant my-app > /run/secrets/my-app-speech-token
|
|
27
|
+
```
|
|
28
|
+
|
|
29
|
+
The service defaults to `127.0.0.1:3092`, with its WebSocket at `/v1/voice` and
|
|
30
|
+
readiness at `/health`. An application's authenticated server proxy consumes the
|
|
31
|
+
token file; neither the token nor signing key belongs in browser configuration.
|
|
32
|
+
The standalone token command does not implement per-application revocation.
|
|
33
|
+
Hosts needing grant rotation/revocation supply their existing grant policy through
|
|
34
|
+
`runVoiceCli(argv, { createDaemon, recognitionVocabulary })` and the generic daemon's authorization seam.
|
|
35
|
+
|
|
36
|
+
Preparation verifies archive hashes, stages exact model files/configuration,
|
|
37
|
+
records file hashes and atomically replaces the model directory. Repeating the
|
|
38
|
+
same preparation reuses a verified pack. Stop the service before replacing its
|
|
39
|
+
pack, verify, then restart. Normal serving loads models without rehashing them on
|
|
40
|
+
every restart. `--download-cache PATH` and `--retain-downloads 1` permit offline
|
|
41
|
+
reuse. `--hotwords-file FILE` replaces the prepared recognizer hotwords file.
|
|
42
|
+
|
|
43
|
+
## Models and voices
|
|
44
|
+
|
|
45
|
+
`--pack cori` (the default) uses the existing single Cori voice and English
|
|
46
|
+
streaming Zipformer recognizer. `--pack piper` keeps that recognizer and offers
|
|
47
|
+
five Piper medium voices at 22,050 Hz:
|
|
48
|
+
|
|
49
|
+
| Voice | Sound |
|
|
50
|
+
| --- | --- |
|
|
51
|
+
| Cori (default) | Female, British English |
|
|
52
|
+
| Alba | Female, Scottish English |
|
|
53
|
+
| Joe | Male, American English |
|
|
54
|
+
| Bryce | Male, American English |
|
|
55
|
+
| Northern English | Male, British English |
|
|
56
|
+
|
|
57
|
+
Each Piper voice has a separate model. The service checks their metadata one at
|
|
58
|
+
a time, then keeps only the default model loaded. Its existing queue synthesizes
|
|
59
|
+
one phrase at a time in one child process. When a request selects a different
|
|
60
|
+
model, that process exits before its replacement loads; this releases the old
|
|
61
|
+
model and native inference caches. The recognition process and audio connection
|
|
62
|
+
stay running. The first reply after a change includes model-loading time;
|
|
63
|
+
subsequent replies reuse it. More choices add disk space without keeping every
|
|
64
|
+
model in RAM. Measure the whole application's memory budget before installation.
|
|
65
|
+
The worker serves the shared queue, so requests from different conversations can
|
|
66
|
+
also require a model change.
|
|
67
|
+
|
|
68
|
+
For embedded use, `await createSherpaSpeechEngine(options)` validates the models
|
|
69
|
+
and returns the engine. Call `await engine.close()` when its owner shuts down.
|
|
70
|
+
The supplied CLI handles both startup failure and shutdown. Cancellation kills
|
|
71
|
+
unfinished native synthesis; the next request recreates that worker if needed.
|
|
72
|
+
|
|
73
|
+
`--pack kokoro` uses Kokoro 1.0 with named female
|
|
74
|
+
Heart, Bella and Emma voices, and male Michael, Adam, George and Daniel voices.
|
|
75
|
+
The service advertises these IDs/labels; the browser selects a speaker per reply.
|
|
76
|
+
Model weights are not included in the npm artifact.
|
|
77
|
+
|
|
78
|
+
`--pack kitten` offers **Bella, Luna, Rosie and Kiki** (female) and
|
|
79
|
+
**Jasper, Bruno, Hugo and Leo** (male) from Kitten TTS Micro 0.8. All eight share
|
|
80
|
+
one model and retain its native 24,000 Hz output. Bella uses a 1.15× speech speed;
|
|
81
|
+
the other voices use normal speed. Kitten does not identify British or Australian
|
|
82
|
+
accents in its voice catalogue.
|
|
83
|
+
`--pack piper-kitten` combines the five Piper choices with those eight Kitten voices,
|
|
84
|
+
listing Kitten first and keeping Cori as the default. It uses Sherpa's resampler inside the synthesis worker
|
|
85
|
+
to deliver every voice at 22,050 Hz. Selecting another engine does not replace the
|
|
86
|
+
audio connection. Only one synthesis model is resident, including when switching
|
|
87
|
+
between Piper and Kitten. Measure Kitten's latency on the target CPU before
|
|
88
|
+
choosing it as a default.
|
|
89
|
+
|
|
90
|
+
Operators may edit a separate speech configuration JSON and pass `serve --config
|
|
91
|
+
FILE` (or `JSKIT_VOICE_CONFIG`). Without an override, serving reads the prepared
|
|
92
|
+
pack's `speech.json`. Native `recognizer` and `synthesizer` configuration objects
|
|
93
|
+
are forwarded to Sherpa after replacing `${MODELS_ROOT}` in paths. Other fields:
|
|
94
|
+
|
|
95
|
+
```json
|
|
96
|
+
{
|
|
97
|
+
"voices": [
|
|
98
|
+
{ "id": "am_michael", "label": "Michael · male · American English", "language": "en-US", "speakerId": 16 },
|
|
99
|
+
{ "id": "bf_emma", "label": "Emma · female · British English", "language": "en-GB", "speakerId": 21 }
|
|
100
|
+
],
|
|
101
|
+
"defaultVoice": "am_michael",
|
|
102
|
+
"recognizerTailPaddingSeconds": 1,
|
|
103
|
+
"recognizerMinimumRms": 0.001,
|
|
104
|
+
"recognitionVocabulary": { "acme labs": "Acme Labs" }
|
|
105
|
+
}
|
|
106
|
+
```
|
|
107
|
+
|
|
108
|
+
`recognitionVocabulary` maps literal recognized phrases to application spellings.
|
|
109
|
+
It applies only when making all-uppercase recognition readable; mixed-case text
|
|
110
|
+
is left intact. The CLI host may supply default vocabulary through its options;
|
|
111
|
+
an explicit speech configuration vocabulary replaces those defaults. This does
|
|
112
|
+
not change the recognizer's hotwords or load another model.
|
|
113
|
+
|
|
114
|
+
Each voice may set `speed` between `0.5` and `2` (default `1`). For example,
|
|
115
|
+
`"speed": 1.15` increases its synthesis pace by about 15% without changing the
|
|
116
|
+
playback sample rate. The setting belongs to that voice; it does not carry over
|
|
117
|
+
when another voice is selected.
|
|
118
|
+
|
|
119
|
+
Merge this example into the Kokoro pack's configuration, retaining its synthesizer
|
|
120
|
+
model paths. Speaker IDs are specific to the selected model. Unknown IDs and
|
|
121
|
+
invalid/duplicate catalogue entries fail explicitly. Voices do not change a
|
|
122
|
+
conversation's reasoning model. Tail padding supplies recognizer lookahead when a
|
|
123
|
+
recording ends immediately after a word; it does not wait for wall-clock silence.
|
|
124
|
+
The RMS floor prevents silent input from starting a recognition utterance. After
|
|
125
|
+
an audible onset, all frames reach the recognizer until the utterance ends. Its
|
|
126
|
+
default is 0.001 (−60 dBFS); configure a positive value up to 0.1 for the input
|
|
127
|
+
device. This is an input-level check, not a classifier for speech versus noise.
|
|
128
|
+
|
|
129
|
+
For separate models, use `synthesizers: { modelId: nativeConfiguration }` instead
|
|
130
|
+
of `synthesizer`, and give each voice a matching `modelId`. Speaker IDs are checked
|
|
131
|
+
against that voice's model. Set `outputSampleRate` (8,000–48,000 Hz) to normalize
|
|
132
|
+
different native rates inside the worker. Without it, all models must use the same
|
|
133
|
+
output sample rate for the shared audio connection. `models/voice-models-piper.json` contains the complete
|
|
134
|
+
five-voice example. The service selects a model per synthesis request; changing a
|
|
135
|
+
voice does not redirect an in-flight reply or another conversation.
|
|
136
|
+
|
|
137
|
+
For other models, `prepare --sources-file FILE` accepts the source manifest shape
|
|
138
|
+
shown in `models/voice-models-piper.json`: recognition and synthesis archives, pinned
|
|
139
|
+
SHA-256 values, explicit source/destination file mappings and native configuration.
|
|
140
|
+
Paths in mappings must stay inside their extracted/staged roots. Model compatibility
|
|
141
|
+
still needs native inference validation; a matching hash is only integrity proof.
|
|
142
|
+
|
|
143
|
+
Thread counts, recognition paths, connection and synthesis-queue limits are
|
|
144
|
+
explicit CLI/environment options (see the CLI). Configuration objects can also
|
|
145
|
+
set native thread counts. Benchmark both first audio latency and sustained audio
|
|
146
|
+
production on the deployment machine before selecting a pack. A higher quality
|
|
147
|
+
voice can consume more CPU and respond more slowly.
|
|
148
|
+
|
|
149
|
+
## Upstream terms and evidence
|
|
150
|
+
|
|
151
|
+
Models and phonemizer data have their own terms, separate from this package:
|
|
152
|
+
|
|
153
|
+
- [Zipformer English model](https://huggingface.co/csukuangfj/sherpa-onnx-streaming-zipformer-en-2023-06-26) declares Apache-2.0.
|
|
154
|
+
- [Cori model card](https://huggingface.co/rhasspy/piper-voices/blob/main/en/en_GB/cori/medium/MODEL_CARD) describes the LibriVox public-domain dataset; the Piper repository declares MIT.
|
|
155
|
+
- [Alba speech corpus](https://doi.org/10.7488/ds/2506): Valentini-Botinhao, Cassia; Yamagishi, Junichi (2019), University of Edinburgh, CC BY 4.0. The Piper model was trained from this corpus; preparation preserves its source notice.
|
|
156
|
+
- [Joe model card](https://huggingface.co/rhasspy/piper-voices/blob/main/en/en_US/joe/medium/MODEL_CARD) declares CC0 source recordings. [Bryce's model card](https://huggingface.co/rhasspy/piper-voices/blob/main/en/en_US/bryce/medium/MODEL_CARD) declares public-domain recordings supplied by the speaker.
|
|
157
|
+
- [Northern English Male model card](https://huggingface.co/rhasspy/piper-voices/blob/main/en/en_GB/northern_english_male/medium/MODEL_CARD) identifies OpenSLR 83 regional UK recordings under CC BY-SA 4.0. Preparation retains the model card.
|
|
158
|
+
- [Kokoro model](https://huggingface.co/hexgrad/Kokoro-82M) declares Apache-2.0. Preparation retains the downloaded model's LICENSE.
|
|
159
|
+
- [Kitten TTS Micro 0.8](https://huggingface.co/KittenML/kitten-tts-micro-0.8) declares Apache-2.0. Preparation retains its LICENSE and Sherpa conversion notes. Bella is speaker 1 and Jasper is speaker 0 in Sherpa's prepared voice table.
|
|
160
|
+
- The archives also contain [eSpeak NG data](https://github.com/espeak-ng/espeak-ng/blob/master/COPYING), with GPL terms. Preserve applicable upstream notices when distributing a prepared pack.
|
|
161
|
+
|
|
162
|
+
`models/` records pinned source archives; `sources.json` and `voice-models.json`
|
|
163
|
+
record what was installed. The optional native test synthesizes and recognizes
|
|
164
|
+
real audio, including male/female cases and cancellation:
|
|
165
|
+
|
|
166
|
+
```sh
|
|
167
|
+
JSKIT_VOICE_TEST_MODELS=/srv/speech/models npm test --workspace packages/assistant-voice-sherpa -- test/nativeSpeech.test.js
|
|
168
|
+
```
|
|
169
|
+
|
|
170
|
+
Physical microphone, speaker, echo and interruption acceptance is still required
|
|
171
|
+
for a full duplex application. Generated-audio round trips cannot establish it.
|
|
@@ -0,0 +1,500 @@
|
|
|
1
|
+
<blk> 0.0
|
|
2
|
+
<sos/eos> 0.0
|
|
3
|
+
<unk> 0.0
|
|
4
|
+
S -3.237640857696533
|
|
5
|
+
▁THE -3.3911421298980713
|
|
6
|
+
▁A -3.912346839904785
|
|
7
|
+
T -3.9788036346435547
|
|
8
|
+
▁AND -4.041482925415039
|
|
9
|
+
ED -4.068052768707275
|
|
10
|
+
▁OF -4.111221790313721
|
|
11
|
+
▁TO -4.148194789886475
|
|
12
|
+
E -4.159254550933838
|
|
13
|
+
D -4.198842525482178
|
|
14
|
+
N -4.289172649383545
|
|
15
|
+
ING -4.401233673095703
|
|
16
|
+
▁IN -4.452512264251709
|
|
17
|
+
Y -4.527078151702881
|
|
18
|
+
M -4.579193592071533
|
|
19
|
+
C -4.626455307006836
|
|
20
|
+
▁I -4.682346820831299
|
|
21
|
+
A -4.6846842765808105
|
|
22
|
+
P -4.696129322052002
|
|
23
|
+
▁HE -4.7119832038879395
|
|
24
|
+
R -4.821435451507568
|
|
25
|
+
O -4.838934898376465
|
|
26
|
+
L -4.880065441131592
|
|
27
|
+
RE -4.8860273361206055
|
|
28
|
+
I -4.903520107269287
|
|
29
|
+
U -4.911571025848389
|
|
30
|
+
ER -4.954335689544678
|
|
31
|
+
▁IT -4.997008323669434
|
|
32
|
+
LY -5.001551628112793
|
|
33
|
+
▁THAT -5.0079545974731445
|
|
34
|
+
▁WAS -5.046072006225586
|
|
35
|
+
▁ -5.05485725402832
|
|
36
|
+
▁S -5.0588836669921875
|
|
37
|
+
AR -5.101287841796875
|
|
38
|
+
▁BE -5.145685195922852
|
|
39
|
+
F -5.181809425354004
|
|
40
|
+
▁C -5.188362121582031
|
|
41
|
+
IN -5.200289249420166
|
|
42
|
+
B -5.208587646484375
|
|
43
|
+
▁FOR -5.234852313995361
|
|
44
|
+
OR -5.269742488861084
|
|
45
|
+
LE -5.270202159881592
|
|
46
|
+
' -5.270315170288086
|
|
47
|
+
▁HIS -5.28671932220459
|
|
48
|
+
▁YOU -5.34528923034668
|
|
49
|
+
AL -5.353833198547363
|
|
50
|
+
▁RE -5.358627796173096
|
|
51
|
+
V -5.368106365203857
|
|
52
|
+
▁B -5.369124412536621
|
|
53
|
+
G -5.393369674682617
|
|
54
|
+
RI -5.41680908203125
|
|
55
|
+
▁E -5.418990135192871
|
|
56
|
+
▁WITH -5.42680025100708
|
|
57
|
+
▁T -5.491506576538086
|
|
58
|
+
▁AS -5.501863956451416
|
|
59
|
+
LL -5.506050109863281
|
|
60
|
+
▁P -5.5140910148620605
|
|
61
|
+
▁HER -5.523913860321045
|
|
62
|
+
ST -5.528979778289795
|
|
63
|
+
▁HAD -5.5388503074646
|
|
64
|
+
▁SO -5.563423156738281
|
|
65
|
+
▁F -5.564723968505859
|
|
66
|
+
W -5.571264266967773
|
|
67
|
+
CE -5.612353324890137
|
|
68
|
+
▁IS -5.6334099769592285
|
|
69
|
+
ND -5.636771202087402
|
|
70
|
+
▁NOT -5.646847248077393
|
|
71
|
+
TH -5.6582183837890625
|
|
72
|
+
▁BUT -5.658850193023682
|
|
73
|
+
EN -5.671176433563232
|
|
74
|
+
▁SHE -5.674525737762451
|
|
75
|
+
▁ON -5.675063610076904
|
|
76
|
+
VE -5.678800582885742
|
|
77
|
+
ON -5.681077480316162
|
|
78
|
+
SE -5.681278228759766
|
|
79
|
+
▁DE -5.684993743896484
|
|
80
|
+
UR -5.702572345733643
|
|
81
|
+
▁G -5.707157611846924
|
|
82
|
+
CH -5.715974807739258
|
|
83
|
+
K -5.731172561645508
|
|
84
|
+
TER -5.7358808517456055
|
|
85
|
+
▁AT -5.74332857131958
|
|
86
|
+
IT -5.752814769744873
|
|
87
|
+
▁ME -5.759988784790039
|
|
88
|
+
RO -5.787921905517578
|
|
89
|
+
NE -5.812368392944336
|
|
90
|
+
RA -5.847346305847168
|
|
91
|
+
ES -5.869145393371582
|
|
92
|
+
IL -5.916404724121094
|
|
93
|
+
NG -5.955217361450195
|
|
94
|
+
IC -5.956216335296631
|
|
95
|
+
▁NO -5.958745002746582
|
|
96
|
+
▁HIM -5.96319580078125
|
|
97
|
+
ENT -5.965415954589844
|
|
98
|
+
IR -5.9872660636901855
|
|
99
|
+
▁WE -6.002829074859619
|
|
100
|
+
H -6.005046367645264
|
|
101
|
+
▁DO -6.016775131225586
|
|
102
|
+
▁ALL -6.021503925323486
|
|
103
|
+
▁HAVE -6.041546821594238
|
|
104
|
+
LO -6.041745662689209
|
|
105
|
+
▁BY -6.075400352478027
|
|
106
|
+
▁MY -6.076292991638184
|
|
107
|
+
▁MO -6.076861381530762
|
|
108
|
+
▁THIS -6.077490329742432
|
|
109
|
+
LA -6.079364776611328
|
|
110
|
+
▁ST -6.100840091705322
|
|
111
|
+
▁WHICH -6.1011643409729
|
|
112
|
+
▁CON -6.115716934204102
|
|
113
|
+
▁THEY -6.133782863616943
|
|
114
|
+
CK -6.134190559387207
|
|
115
|
+
TE -6.148074626922607
|
|
116
|
+
▁SAID -6.1597394943237305
|
|
117
|
+
▁FROM -6.16114616394043
|
|
118
|
+
▁GO -6.164953708648682
|
|
119
|
+
▁WHO -6.172391414642334
|
|
120
|
+
▁TH -6.1845221519470215
|
|
121
|
+
▁OR -6.188831329345703
|
|
122
|
+
▁D -6.1897478103637695
|
|
123
|
+
▁W -6.1943888664245605
|
|
124
|
+
VER -6.204941272735596
|
|
125
|
+
LI -6.220948219299316
|
|
126
|
+
▁SE -6.229703426361084
|
|
127
|
+
▁ONE -6.230255126953125
|
|
128
|
+
▁CA -6.233973979949951
|
|
129
|
+
▁AN -6.242618560791016
|
|
130
|
+
▁LA -6.247861385345459
|
|
131
|
+
▁WERE -6.248575687408447
|
|
132
|
+
EL -6.27099609375
|
|
133
|
+
▁HA -6.283683776855469
|
|
134
|
+
▁MAN -6.2989583015441895
|
|
135
|
+
▁FA -6.311679363250732
|
|
136
|
+
▁EX -6.313636302947998
|
|
137
|
+
AD -6.315258979797363
|
|
138
|
+
▁SU -6.340909957885742
|
|
139
|
+
RY -6.354572296142578
|
|
140
|
+
▁MI -6.370210647583008
|
|
141
|
+
AT -6.391071796417236
|
|
142
|
+
▁BO -6.3981523513793945
|
|
143
|
+
▁WHEN -6.406372547149658
|
|
144
|
+
AN -6.418525695800781
|
|
145
|
+
THER -6.4211506843566895
|
|
146
|
+
PP -6.437921524047852
|
|
147
|
+
ATION -6.441051006317139
|
|
148
|
+
▁FI -6.457613945007324
|
|
149
|
+
▁WOULD -6.472571849822998
|
|
150
|
+
▁PRO -6.472743511199951
|
|
151
|
+
OW -6.480467319488525
|
|
152
|
+
ET -6.48289155960083
|
|
153
|
+
▁O -6.487629413604736
|
|
154
|
+
▁THERE -6.488083839416504
|
|
155
|
+
▁HO -6.488405704498291
|
|
156
|
+
ION -6.488462448120117
|
|
157
|
+
▁WHAT -6.491586685180664
|
|
158
|
+
▁FE -6.4979729652404785
|
|
159
|
+
▁PA -6.500291347503662
|
|
160
|
+
US -6.500805854797363
|
|
161
|
+
MENT -6.503949165344238
|
|
162
|
+
▁MA -6.505156993865967
|
|
163
|
+
UT -6.538065433502197
|
|
164
|
+
▁OUT -6.54595947265625
|
|
165
|
+
▁THEIR -6.548971652984619
|
|
166
|
+
▁IF -6.550684452056885
|
|
167
|
+
▁LI -6.554531574249268
|
|
168
|
+
▁K -6.563741207122803
|
|
169
|
+
▁WILL -6.566903591156006
|
|
170
|
+
▁ARE -6.574769020080566
|
|
171
|
+
ID -6.575021266937256
|
|
172
|
+
▁RO -6.575534820556641
|
|
173
|
+
DE -6.581881999969482
|
|
174
|
+
TION -6.595175743103027
|
|
175
|
+
▁WA -6.5956645011901855
|
|
176
|
+
PE -6.601837635040283
|
|
177
|
+
▁UP -6.603113174438477
|
|
178
|
+
▁SP -6.611117362976074
|
|
179
|
+
▁PO -6.614088535308838
|
|
180
|
+
IGHT -6.637707710266113
|
|
181
|
+
▁UN -6.641831398010254
|
|
182
|
+
RU -6.642515182495117
|
|
183
|
+
▁LO -6.646393775939941
|
|
184
|
+
AS -6.648888111114502
|
|
185
|
+
OL -6.654422283172607
|
|
186
|
+
▁LE -6.663815975189209
|
|
187
|
+
▁BEEN -6.678056716918945
|
|
188
|
+
▁SH -6.678354263305664
|
|
189
|
+
▁RA -6.680027961730957
|
|
190
|
+
▁SEE -6.69546365737915
|
|
191
|
+
KE -6.708260536193848
|
|
192
|
+
UL -6.716123580932617
|
|
193
|
+
TED -6.716958522796631
|
|
194
|
+
▁SA -6.729671955108643
|
|
195
|
+
UN -6.731771945953369
|
|
196
|
+
UND -6.733950614929199
|
|
197
|
+
ANT -6.734233379364014
|
|
198
|
+
▁NE -6.7430009841918945
|
|
199
|
+
IS -6.754036903381348
|
|
200
|
+
▁THEM -6.757429599761963
|
|
201
|
+
CI -6.769925594329834
|
|
202
|
+
GE -6.770693778991699
|
|
203
|
+
▁COULD -6.775028228759766
|
|
204
|
+
▁DIS -6.7813544273376465
|
|
205
|
+
OM -6.793661117553711
|
|
206
|
+
ISH -6.803374290466309
|
|
207
|
+
HE -6.807046413421631
|
|
208
|
+
EST -6.810579299926758
|
|
209
|
+
▁SOME -6.817039966583252
|
|
210
|
+
ENCE -6.822673320770264
|
|
211
|
+
ITY -6.835846424102783
|
|
212
|
+
IVE -6.840222358703613
|
|
213
|
+
▁US -6.840911865234375
|
|
214
|
+
▁MORE -6.844557285308838
|
|
215
|
+
▁EN -6.849472522735596
|
|
216
|
+
ARD -6.851588249206543
|
|
217
|
+
ATE -6.8606672286987305
|
|
218
|
+
▁YOUR -6.861882209777832
|
|
219
|
+
▁INTO -6.862520694732666
|
|
220
|
+
▁KNOW -6.867526531219482
|
|
221
|
+
▁CO -6.86960506439209
|
|
222
|
+
ANCE -6.872697830200195
|
|
223
|
+
▁TIME -6.875100612640381
|
|
224
|
+
▁WI -6.883917331695557
|
|
225
|
+
▁YE -6.887696743011475
|
|
226
|
+
AGE -6.895497798919678
|
|
227
|
+
▁NOW -6.903868198394775
|
|
228
|
+
TI -6.904016017913818
|
|
229
|
+
FF -6.906183242797852
|
|
230
|
+
ABLE -6.909174919128418
|
|
231
|
+
▁VERY -6.922069072723389
|
|
232
|
+
▁LIKE -6.928692817687988
|
|
233
|
+
AM -6.946870803833008
|
|
234
|
+
HI -6.949864864349365
|
|
235
|
+
Z -6.951096534729004
|
|
236
|
+
▁OTHER -6.964641094207764
|
|
237
|
+
▁THAN -6.97308349609375
|
|
238
|
+
▁LITTLE -6.976223468780518
|
|
239
|
+
▁DID -6.981521129608154
|
|
240
|
+
▁LOOK -6.9846696853637695
|
|
241
|
+
TY -6.9947428703308105
|
|
242
|
+
ERS -6.999180793762207
|
|
243
|
+
▁CAN -7.009943962097168
|
|
244
|
+
▁CHA -7.01063346862793
|
|
245
|
+
▁AR -7.025554656982422
|
|
246
|
+
X -7.026675701141357
|
|
247
|
+
FUL -7.034653186798096
|
|
248
|
+
UGH -7.046075344085693
|
|
249
|
+
▁BA -7.0493693351745605
|
|
250
|
+
▁DAY -7.054299831390381
|
|
251
|
+
▁ABOUT -7.0555419921875
|
|
252
|
+
TEN -7.055879592895508
|
|
253
|
+
IM -7.058297157287598
|
|
254
|
+
▁ANY -7.058923721313477
|
|
255
|
+
▁PRE -7.061312675476074
|
|
256
|
+
▁OVER -7.064944744110107
|
|
257
|
+
IES -7.078317165374756
|
|
258
|
+
NESS -7.08111047744751
|
|
259
|
+
ME -7.0836920738220215
|
|
260
|
+
BLE -7.086453437805176
|
|
261
|
+
▁M -7.090318202972412
|
|
262
|
+
ROW -7.09219217300415
|
|
263
|
+
▁HAS -7.097603797912598
|
|
264
|
+
▁GREAT -7.10443639755249
|
|
265
|
+
▁VI -7.105268955230713
|
|
266
|
+
TA -7.107810020446777
|
|
267
|
+
▁AFTER -7.109515190124512
|
|
268
|
+
PER -7.113203048706055
|
|
269
|
+
▁AGAIN -7.114784240722656
|
|
270
|
+
HO -7.115448474884033
|
|
271
|
+
SH -7.115546226501465
|
|
272
|
+
▁UPON -7.1267476081848145
|
|
273
|
+
▁DI -7.130036354064941
|
|
274
|
+
▁HAND -7.132112979888916
|
|
275
|
+
▁COM -7.132730484008789
|
|
276
|
+
IST -7.134383201599121
|
|
277
|
+
TURE -7.137714385986328
|
|
278
|
+
▁STA -7.146283149719238
|
|
279
|
+
▁THEN -7.1577534675598145
|
|
280
|
+
▁SHOULD -7.158360481262207
|
|
281
|
+
▁GA -7.167736053466797
|
|
282
|
+
OUS -7.169438362121582
|
|
283
|
+
OUR -7.169652462005615
|
|
284
|
+
▁WELL -7.171218395233154
|
|
285
|
+
▁ONLY -7.176281452178955
|
|
286
|
+
MAN -7.178989410400391
|
|
287
|
+
▁GOOD -7.184938430786133
|
|
288
|
+
▁TWO -7.1874613761901855
|
|
289
|
+
▁MAR -7.1875505447387695
|
|
290
|
+
▁SAY -7.196364879608154
|
|
291
|
+
▁HU -7.202139377593994
|
|
292
|
+
TING -7.208182334899902
|
|
293
|
+
▁OUR -7.2181267738342285
|
|
294
|
+
RESS -7.222963333129883
|
|
295
|
+
▁DOWN -7.223188400268555
|
|
296
|
+
IOUS -7.247483730316162
|
|
297
|
+
▁BEFORE -7.248350620269775
|
|
298
|
+
▁DA -7.250795364379883
|
|
299
|
+
▁NA -7.258609294891357
|
|
300
|
+
QUI -7.263876438140869
|
|
301
|
+
▁MADE -7.269183158874512
|
|
302
|
+
▁EVERY -7.269209384918213
|
|
303
|
+
▁OLD -7.277935981750488
|
|
304
|
+
▁EVEN -7.280048370361328
|
|
305
|
+
IG -7.2808051109313965
|
|
306
|
+
▁COME -7.281715393066406
|
|
307
|
+
▁GRA -7.285216808319092
|
|
308
|
+
▁RI -7.291446685791016
|
|
309
|
+
▁LONG -7.293140411376953
|
|
310
|
+
OT -7.296908378601074
|
|
311
|
+
SIDE -7.300649166107178
|
|
312
|
+
WARD -7.311554431915283
|
|
313
|
+
▁FO -7.31390905380249
|
|
314
|
+
▁WHERE -7.314794540405273
|
|
315
|
+
MO -7.319922924041748
|
|
316
|
+
LESS -7.328833103179932
|
|
317
|
+
▁SC -7.3289055824279785
|
|
318
|
+
▁MUST -7.330503463745117
|
|
319
|
+
▁NEVER -7.330777645111084
|
|
320
|
+
▁HOW -7.343449592590332
|
|
321
|
+
▁CAME -7.345986843109131
|
|
322
|
+
▁SUCH -7.347212791442871
|
|
323
|
+
▁RU -7.359615802764893
|
|
324
|
+
▁TAKE -7.359617710113525
|
|
325
|
+
▁WO -7.368528366088867
|
|
326
|
+
▁CAR -7.379446029663086
|
|
327
|
+
UM -7.379636764526367
|
|
328
|
+
AK -7.392054557800293
|
|
329
|
+
▁THINK -7.40938663482666
|
|
330
|
+
▁MUCH -7.409630298614502
|
|
331
|
+
▁MISTER -7.420104503631592
|
|
332
|
+
▁MAY -7.43388557434082
|
|
333
|
+
▁JO -7.443871021270752
|
|
334
|
+
▁WAY -7.445600509643555
|
|
335
|
+
▁COMP -7.455471515655518
|
|
336
|
+
▁THOUGHT -7.456057548522949
|
|
337
|
+
▁STO -7.462839603424072
|
|
338
|
+
▁MEN -7.463774681091309
|
|
339
|
+
▁BACK -7.465264797210693
|
|
340
|
+
▁DON -7.465480327606201
|
|
341
|
+
J -7.466047286987305
|
|
342
|
+
▁LET -7.483911514282227
|
|
343
|
+
▁TRA -7.495802879333496
|
|
344
|
+
▁FIRST -7.496084213256836
|
|
345
|
+
▁JUST -7.497234344482422
|
|
346
|
+
▁VA -7.499629497528076
|
|
347
|
+
▁OWN -7.5111403465271
|
|
348
|
+
▁PLA -7.516643047332764
|
|
349
|
+
▁MAKE -7.521800994873047
|
|
350
|
+
ATED -7.529211521148682
|
|
351
|
+
▁HIMSELF -7.533393383026123
|
|
352
|
+
▁WENT -7.539999485015869
|
|
353
|
+
▁PI -7.557174205780029
|
|
354
|
+
GG -7.559944152832031
|
|
355
|
+
RING -7.560603618621826
|
|
356
|
+
▁DU -7.5634050369262695
|
|
357
|
+
▁MIGHT -7.56861686706543
|
|
358
|
+
▁PART -7.570089817047119
|
|
359
|
+
▁GIVE -7.583937168121338
|
|
360
|
+
▁IMP -7.588171005249023
|
|
361
|
+
▁BU -7.591918468475342
|
|
362
|
+
▁PER -7.595826148986816
|
|
363
|
+
▁PLACE -7.605018138885498
|
|
364
|
+
▁HOUSE -7.609342098236084
|
|
365
|
+
▁THROUGH -7.625210285186768
|
|
366
|
+
IAN -7.628554821014404
|
|
367
|
+
▁SW -7.637818336486816
|
|
368
|
+
▁UNDER -7.643731117248535
|
|
369
|
+
QUE -7.644018173217773
|
|
370
|
+
▁AWAY -7.645115375518799
|
|
371
|
+
▁LOVE -7.6468071937561035
|
|
372
|
+
QUA -7.650307655334473
|
|
373
|
+
▁LIFE -7.660292148590088
|
|
374
|
+
▁GET -7.666475772857666
|
|
375
|
+
▁WITHOUT -7.6741156578063965
|
|
376
|
+
▁PASS -7.682002544403076
|
|
377
|
+
▁TURN -7.692363262176514
|
|
378
|
+
IGN -7.6964945793151855
|
|
379
|
+
▁HEAD -7.696974754333496
|
|
380
|
+
▁MOST -7.705150604248047
|
|
381
|
+
▁THOSE -7.7168049812316895
|
|
382
|
+
▁SHALL -7.717948913574219
|
|
383
|
+
▁EYES -7.718607425689697
|
|
384
|
+
▁COL -7.741212844848633
|
|
385
|
+
▁STILL -7.7420759201049805
|
|
386
|
+
▁NIGHT -7.745421886444092
|
|
387
|
+
▁NOTHING -7.768163681030273
|
|
388
|
+
ITION -7.768242835998535
|
|
389
|
+
HA -7.769926071166992
|
|
390
|
+
▁TELL -7.769962787628174
|
|
391
|
+
▁WORK -7.773743152618408
|
|
392
|
+
▁LAST -7.775449752807617
|
|
393
|
+
▁NEW -7.78277063369751
|
|
394
|
+
▁FACE -7.78385591506958
|
|
395
|
+
▁HI -7.791258811950684
|
|
396
|
+
▁WORD -7.8012003898620605
|
|
397
|
+
▁FOUND -7.80479097366333
|
|
398
|
+
▁COUNT -7.8051581382751465
|
|
399
|
+
▁OB -7.805266380310059
|
|
400
|
+
▁WHILE -7.807648658752441
|
|
401
|
+
▁SHA -7.817192077636719
|
|
402
|
+
▁MEAN -7.83618688583374
|
|
403
|
+
▁SAW -7.836764335632324
|
|
404
|
+
▁PEOPLE -7.837661266326904
|
|
405
|
+
▁FRIEND -7.85650110244751
|
|
406
|
+
▁THREE -7.868136882781982
|
|
407
|
+
▁ROOM -7.875853538513184
|
|
408
|
+
▁SAME -7.887514114379883
|
|
409
|
+
▁THOUGH -7.892477035522461
|
|
410
|
+
▁RIGHT -7.896188259124756
|
|
411
|
+
▁CHILD -7.897783279418945
|
|
412
|
+
▁FATHER -7.903822898864746
|
|
413
|
+
▁ANOTHER -7.905872821807861
|
|
414
|
+
▁HEART -7.912759304046631
|
|
415
|
+
▁WANT -7.922352313995361
|
|
416
|
+
▁TOOK -7.94252872467041
|
|
417
|
+
OOK -7.9485063552856445
|
|
418
|
+
▁LIGHT -7.961446285247803
|
|
419
|
+
▁MISSUS -7.978604316711426
|
|
420
|
+
▁OPEN -7.985271453857422
|
|
421
|
+
▁JU -7.989404678344727
|
|
422
|
+
▁ASKED -7.9909563064575195
|
|
423
|
+
PORT -8.001195907592773
|
|
424
|
+
▁LEFT -8.001866340637207
|
|
425
|
+
▁JA -8.029788970947266
|
|
426
|
+
▁WORLD -8.0399169921875
|
|
427
|
+
▁HOME -8.049389839172363
|
|
428
|
+
▁WHY -8.061796188354492
|
|
429
|
+
▁ALWAYS -8.065637588500977
|
|
430
|
+
▁ANSWER -8.073901176452637
|
|
431
|
+
▁SEEMED -8.087997436523438
|
|
432
|
+
▁SOMETHING -8.089508056640625
|
|
433
|
+
▁GIRL -8.09334945678711
|
|
434
|
+
▁BECAUSE -8.106149673461914
|
|
435
|
+
▁NAME -8.107681274414062
|
|
436
|
+
▁TOLD -8.11745548248291
|
|
437
|
+
▁NI -8.1217622756958
|
|
438
|
+
▁HIGH -8.121779441833496
|
|
439
|
+
IZE -8.138053894042969
|
|
440
|
+
▁WOMAN -8.145215034484863
|
|
441
|
+
▁FOLLOW -8.151243209838867
|
|
442
|
+
▁RETURN -8.171257019042969
|
|
443
|
+
▁KNEW -8.175785064697266
|
|
444
|
+
▁EACH -8.180315971374512
|
|
445
|
+
▁KIND -8.187873840332031
|
|
446
|
+
▁JE -8.19006633758545
|
|
447
|
+
▁ACT -8.200040817260742
|
|
448
|
+
▁LU -8.205354690551758
|
|
449
|
+
▁CERTAIN -8.210489273071289
|
|
450
|
+
▁YEARS -8.227497100830078
|
|
451
|
+
▁QUITE -8.228480339050293
|
|
452
|
+
▁APPEAR -8.234442710876465
|
|
453
|
+
▁BETTER -8.248871803283691
|
|
454
|
+
▁HALF -8.25101089477539
|
|
455
|
+
▁PRESENT -8.251077651977539
|
|
456
|
+
▁PRINCE -8.262996673583984
|
|
457
|
+
SHIP -8.266542434692383
|
|
458
|
+
▁ALSO -8.270049095153809
|
|
459
|
+
▁BEGAN -8.274964332580566
|
|
460
|
+
▁HAVING -8.284735679626465
|
|
461
|
+
▁ENOUGH -8.286083221435547
|
|
462
|
+
▁PERSON -8.28836441040039
|
|
463
|
+
▁LADY -8.294978141784668
|
|
464
|
+
▁WHITE -8.311501502990723
|
|
465
|
+
▁COURSE -8.316527366638184
|
|
466
|
+
▁VOICE -8.319303512573242
|
|
467
|
+
▁SPEAK -8.327986717224121
|
|
468
|
+
▁POWER -8.352112770080566
|
|
469
|
+
▁MORNING -8.355817794799805
|
|
470
|
+
▁BETWEEN -8.356502532958984
|
|
471
|
+
▁AMONG -8.359968185424805
|
|
472
|
+
▁KEEP -8.362163543701172
|
|
473
|
+
▁WALK -8.368597984313965
|
|
474
|
+
▁MATTER -8.370609283447266
|
|
475
|
+
▁TEA -8.378629684448242
|
|
476
|
+
▁BELIEVE -8.37922477722168
|
|
477
|
+
▁SMALL -8.381397247314453
|
|
478
|
+
▁TALK -8.389386177062988
|
|
479
|
+
▁FELT -8.394773483276367
|
|
480
|
+
▁HORSE -8.39834213256836
|
|
481
|
+
▁MYSELF -8.400507926940918
|
|
482
|
+
▁SIX -8.400691032409668
|
|
483
|
+
▁HOWEVER -8.403042793273926
|
|
484
|
+
▁FULL -8.403987884521484
|
|
485
|
+
▁HERSELF -8.407135009765625
|
|
486
|
+
▁POINT -8.412476539611816
|
|
487
|
+
▁STOOD -8.413363456726074
|
|
488
|
+
▁HUNDRED -8.414043426513672
|
|
489
|
+
▁ALMOST -8.42801570892334
|
|
490
|
+
▁SINCE -8.436264991760254
|
|
491
|
+
▁LARGE -8.44206714630127
|
|
492
|
+
▁LEAVE -8.446986198425293
|
|
493
|
+
▁PERHAPS -8.457595825195312
|
|
494
|
+
▁DARK -8.468853950500488
|
|
495
|
+
▁SUDDEN -8.469058990478516
|
|
496
|
+
▁REPLIED -8.475360870361328
|
|
497
|
+
▁ANYTHING -8.48241138458252
|
|
498
|
+
▁WONDER -8.48791790008545
|
|
499
|
+
▁UNTIL -8.489263534545898
|
|
500
|
+
Q -9.788320541381836
|