---
tags:
- espnet
- audio
- automatic-speech-recognition
language: de
datasets:
- speechcatcher
license: mit
---
## Speechcatcher ESPnet streaming ASR model XL for German ASR
### `speechcatcher/speechcatcher_german_espnet_streaming_transformer_26k_train_size_xl_raw_de_bpe1024`
This model was trained by bmilde using speechcatcher recipe in [espnet](https://github.com/speechcatcher-asr/espnet/tree/egs2-speechcatcher-de).
### Demo: How to use the model
Global installation:
```bash
sudo apt-get install portaudio19-dev python3.10-dev ffmpeg
# on mac:
# brew install portaudio ffmpeg
pip3 install git+https://github.com/speechcatcher-asr/speechcatcher
speechcatcher -m de_streaming_transformer_xl mediafile.mp4
# or with a microphone:
speechcatcher -m de_streaming_transformer_xl -l
```
Virtual environment:
```bash
virtualenv -p python3.10 speechcatcher_env
source speechcatcher_env/bin/activate
pip3 install git+https://github.com/speechcatcher-asr/speechcatcher
speechcatcher -m de_streaming_transformer_xl mediafile.mp4
# or with a microphone:
speechcatcher -m de_streaming_transformer_xl -l
```
# RESULTS
Tuda-de-raw: 2.76% CER (without LM)
Tuda-de-raw: 9.65% WER (without LM)
Note: Tuda-de-raw results are based on raw tuda-de test utterances without the normalization step. It may not be directly comparable to regular tuda-de results.
# Speechcatcher training
Speechcatcher models are trained by using Whisper large as a teacher model:
See [speechcatcher-data](https://github.com/speechcatcher-asr/speechcatcher-data) for code and more info on replicating the training process.
# Sponsors
Speechcatcher was gracefully funded by
Media Tech Lab by Media Lab Bayern (@media-tech-lab)
# Citing
```BibTex
@misc{milde2023speechcatcher,
author = {Milde, Benjamin},
title = {Speechcatcher},
year = {2023},
publisher = {GitHub},
journal = {GitHub repository},
howpublished = {\url{https://github.com/speechcatcher-asr/speechcatcher}},
}
```
## ASR config
expand
```
config: conf/train_asr_streaming_transformer_size_l.yaml
print_config: false
log_level: INFO
dry_run: false
iterator_type: sequence
output_dir: exp/asr_train_asr_streaming_transformer_size_l_raw_de_bpe1024
ngpu: 1
seed: 0
num_workers: 1
num_att_plot: 0
dist_backend: nccl
dist_init_method: env://
dist_world_size: 4
dist_rank: 0
local_rank: 0
dist_master_addr: localhost
dist_master_port: 55625
dist_launcher: null
multiprocessing_distributed: true
unused_parameters: false
sharded_ddp: false
cudnn_enabled: true
cudnn_benchmark: false
cudnn_deterministic: true
collect_stats: false
write_collected_feats: false
max_epoch: 20
patience: 3
val_scheduler_criterion:
- valid
- acc
early_stopping_criterion:
- valid
- acc
- max
best_model_criterion:
- - valid
- acc
- max
keep_nbest_models: 10
nbest_averaging_interval: 0
grad_clip: 5
grad_clip_type: 2.0
grad_noise: false
accum_grad: 1
no_forward_run: false
resume: true
train_dtype: float32
use_amp: false
log_interval: null
use_matplotlib: true
use_tensorboard: true
create_graph_in_tensorboard: false
use_wandb: false
wandb_project: null
wandb_id: null
wandb_entity: null
wandb_name: null
wandb_model_log_interval: -1
detect_anomaly: false
pretrain_path: null
init_param: []
ignore_init_mismatch: false
freeze_param: []
num_iters_per_epoch: null
batch_size: 64
valid_batch_size: null
batch_bins: 1000000
valid_batch_bins: null
train_shape_file:
- exp/asr_stats_raw_de_bpe1024/train/speech_shape
- exp/asr_stats_raw_de_bpe1024/train/text_shape.bpe
valid_shape_file:
- exp/asr_stats_raw_de_bpe1024/valid/speech_shape
- exp/asr_stats_raw_de_bpe1024/valid/text_shape.bpe
batch_type: folded
valid_batch_type: null
fold_length:
- 80000
- 150
sort_in_batch: descending
sort_batch: descending
multiple_iterator: false
chunk_length: 500
chunk_shift_ratio: 0.5
num_cache_chunks: 1024
train_data_path_and_name_and_type:
- - dump/raw/train/wav.scp
- speech
- sound
- - dump/raw/train/text
- text
- text
valid_data_path_and_name_and_type:
- - dump/raw/dev/wav.scp
- speech
- sound
- - dump/raw/dev/text
- text
- text
allow_variable_data_keys: false
max_cache_size: 0.0
max_cache_fd: 32
valid_max_cache_size: null
exclude_weight_decay: false
exclude_weight_decay_conf: {}
optim: adam
optim_conf:
lr: 0.001
scheduler: warmuplr
scheduler_conf:
warmup_steps: 25000
token_list:
-
-
- ','
- .
- t
- ▁
- e
- en
- s
- n
- ▁ich
- ▁das
- ▁und
- ▁die
- er
- ▁ist
- ▁auch
- ▁so
- st
- ▁der
- ▁nicht
- ▁es
- ▁ein
- r
- ▁in
- f
- ▁dann
- ▁ja
- d
- ▁da
- g
- h
- m
- o
- u
- b
- ▁wir
- ▁zu
- ▁du
- ▁ge
- ▁Und
- i
- a
- ▁mit
- ▁den
- in
- ▁man
- l
- ▁auf
- ▁dass
- sch
- ▁jetzt
- '?'
- ge
- ▁was
- ▁er
- ▁Ja
- ▁hat
- '-'
- p
- ▁war
- ▁eine
- ▁F
- ▁aber
- ▁mal
- ▁oder
- y
- ▁noch
- te
- ung
- ▁haben
- ▁Ich
- ▁be
- ▁Das
- ▁wie
- ä
- ▁an
- ▁habe
- k
- ▁von
- ▁sich
- ▁K
- al
- ▁wenn
- la
- ▁schon
- ig
- ra
- lich
- re
- de
- ch
- ▁für
- it
- ▁Also
- w
- ▁A
- es
- ▁sind
- ▁ver
- le
- or
- ▁sie
- ▁B
- ü
- ▁also
- ▁ganz
- ▁T
- ▁im
- ▁dem
- ter
- an
- ck
- ▁St
- ▁aus
- ▁G
- ▁kann
- ▁bei
- ▁halt
- ▁H
- el
- ▁immer
- z
- ▁einfach
- ▁P
- ö
- ▁S
- ▁weil
- ▁mir
- se
- ▁f
- ut
- ten
- ▁wo
- ▁Sch
- us
- ▁vor
- ur
- ▁sehr
- ri
- kt
- ing
- ▁E
- il
- ▁gut
- ▁mich
- ▁Aber
- 'on'
- und
- cht
- ▁als
- den
- ar
- ie
- um
- ▁uns
- ste
- ▁Da
- hr
- ▁über
- be
- ▁einen
- ▁Be
- ▁ihr
- is
- ▁wieder
- ▁glaube
- ▁Ge
- at
- ▁irgendwie
- li
- ▁nur
- we
- ro
- ▁bisschen
- he
- ▁mehr
- ▁M
- tz
- ▁muss
- gen
- ▁sagen
- ben
- ▁wirklich
- ▁alle
- nd
- ▁wird
- ▁gibt
- ▁um
- ▁m
- ▁natürlich
- ▁viel
- me
- nt
- et
- ▁diese
- ▁U
- '0'
- ▁sein
- ▁nach
- ▁hier
- ▁meine
- ern
- lo
- ion
- ▁eigentlich
- ▁O
- ▁machen
- ▁bin
- ▁So
- ll
- ▁hast
- ▁weiß
- ▁Re
- c
- ▁I
- ▁sch
- ▁C
- ▁vielleicht
- iert
- ach
- ▁b
- ne
- x
- ze
- rei
- ru
- ma
- ▁zum
- ▁finde
- ß
- ▁N
- ▁Die
- rt
- ich
- ▁Ma
- uch
- ▁eben
- rü
- ▁Ver
- ein
- ▁In
- R
- ieren
- ▁Ha
- ssen
- ft
- chen
- am
- di
- der
- hl
- ▁Es
- ▁gesagt
- zu
- ▁ne
- ▁An
- ▁k
- ▁1
- ▁am
- hn
- ▁gerade
- pp
- her
- ▁alles
- nen
- ▁geht
- ▁genau
- ha
- ▁Jahr
- ▁re
- ▁werden
- ▁w
- ▁Z
- isch
- ▁p
- ▁Er
- ke
- ▁Wir
- au
- mm
- ik
- ▁mein
- ▁dir
- ▁einem
- un
- ▁würde
- ▁We
- ▁zwei
- v
- ▁doch
- ▁keine
- ▁erst
- na
- and
- ▁gar
- ▁hin
- ▁durch
- ▁V
- kommen
- ell
- ul
- end
- ▁können
- j
- fe
- ▁richtig
- ff
- ▁Me
- ▁andere
- lie
- '...'
- wi
- ol
- art
- ▁Leute
- ▁Zeit
- ▁Ein
- ran
- ner
- ▁ab
- nk
- ation
- ▁viele
- ▁g
- S
- rie
- ▁ob
- im
- ver
- ür
- rk
- ▁einer
- men
- ▁ent
- iv
- lei
- ▁gemacht
- sp
- ▁hatte
- ▁weiter
- sten
- che
- ang
- all
- ir
- hör
- ▁Was
- aus
- ier
- ▁Ne
- ▁Li
- ▁hab
- ass
- L
- igen
- zi
- ungen
- ▁Spiel
- ▁will
- ▁unter
- ag
- ▁macht
- ber
- ▁Sp
- zen
- ▁denn
- ken
- ▁des
- ▁Ka
- lle
- id
- sen
- ▁dich
- ▁st
- ▁Du
- ▁kommt
- spiel
- ▁Fall
- ▁Man
- ▁Se
- ▁W
- ▁dieser
- ▁Ko
- ga
- ▁De
- ▁groß
- ▁Le
- ▁schön
- ▁La
- ▁jeden
- ▁D
- ▁Genau
- gt
- ▁dieses
- ungs
- ▁J
- pro
- ▁Co
- ▁Beispiel
- ▁heißt
- ▁s
- ist
- rä
- ho
- ▁damit
- ▁Wo
- ▁unsere
- ▁le
- ert
- '5'
- ni
- tt
- gel
- ▁her
- ve
- ▁sondern
- mp
- reich
- ▁Sa
- ''''
- ▁lang
- ▁rein
- ▁neu
- ▁sagt
- ▁tatsächlich
- ▁kein
- är
- nehmen
- ▁bis
- elt
- ad
- teil
- ▁euch
- ta
- ▁a
- ▁anderen
- ▁raus
- op
- ▁Der
- ige
- arbeit
- ▁Film
- ▁Ba
- ▁heute
- ▁wäre
- ▁nochmal
- ▁ange
- ▁Sie
- ick
- ▁of
- ler
- ▁un
- ische
- weise
- lä
- kl
- ▁Na
- iß
- wa
- ▁wer
- ▁Ding
- ▁okay
- ▁Ra
- halt
- ▁we
- ▁Pa
- ▁Thema
- heit
- ▁ko
- ▁Dann
- ▁diesen
- schaft
- ▁möchte
- ▁hätte
- lu
- ▁Al
- bar
- ▁Tag
- mo
- ▁Wie
- ▁waren
- ▁sp
- ▁wurde
- ▁Auf
- ce
- ▁Frage
- ▁kannst
- wo
- ▁Mi
- ▁deine
- ▁To
- mi
- ▁dazu
- äng
- ▁bist
- ischen
- ▁Mo
- ▁ihn
- 'no'
- zieh
- ▁Ab
- ▁kommen
- ▁Menschen
- anz
- ▁Wenn
- ▁ha
- ▁Vor
- ▁Ro
- stell
- ▁Zu
- ▁je
- rau
- eln
- ab
- hin
- ka
- schau
- ▁Pro
- ger
- P
- ▁Bo
- ▁gerne
- ko
- nis
- ▁drei
- ▁gleich
- ld
- ▁klar
- ack
- ▁Aus
- ün
- ▁nie
- A
- ▁tr
- ▁seine
- ▁Mit
- geben
- ▁soll
- '4'
- ▁diesem
- lau
- ▁müssen
- ▁kleine
- ▁kurz
- mmer
- ment
- stellen
- ▁Wa
- ▁Podcast
- ▁Wi
- ▁the
- ▁Woche
- ▁guck
- ▁quasi
- ▁Ho
- mal
- ▁sei
- ▁Po
- krieg
- aff
- ▁nächste
- itz
- ▁20
- tag
- '9'
- ▁Ende
- richt
- uck
- ör
- ▁2
- dem
- mpf
- vi
- 'off'
- ▁Leben
- ▁wichtig
- ▁gesehen
- ▁gehen
- ress
- ▁sag
- M
- ▁echt
- ▁etwas
- stand
- zähl
- führ
- T
- ▁wenig
- ▁zusammen
- ▁paar
- ▁Di
- ▁einmal
- bo
- ▁sehen
- ▁Sachen
- ▁Kon
- bi
- ▁dabei
- gend
- pass
- ic
- ▁könnte
- ▁Weil
- zeit
- ▁denke
- F
- ▁Folge
- man
- ▁wollte
- kauf
- ▁weg
- ▁3
- ▁selbst
- '1'
- hol
- co
- ▁wollen
- bau
- '2'
- B
- ▁wahrscheinlich
- ank
- ▁Mal
- ▁letzten
- fahren
- ▁vom
- ▁Do
- hi
- ▁eher
- D
- ▁selber
- ord
- ▁super
- ▁musst
- ▁drauf
- ▁jemand
- '8'
- ▁gegen
- ▁überhaupt
- ▁The
- ▁Okay
- ▁beim
- ▁sage
- pa
- ▁dafür
- vor
- ▁Frau
- ▁hatten
- ▁drin
- '6'
- ▁sozusagen
- iz
- ▁fand
- ▁Tra
- folg
- ▁Nach
- ▁tun
- ▁dein
- ität
- C
- ▁Oder
- ▁zurück
- ▁Nein
- po
- ▁cool
- ▁sowas
- ▁sieht
- gehen
- schi
- ▁Gott
- ▁schnell
- form
- ▁ihm
- ▁besser
- ▁gab
- wä
- ▁äh
- ▁Kinder
- änder
- ▁sollte
- ▁Jo
- ▁voll
- ▁War
- ▁kenne
- ▁zwar
- ▁total
- ▁welche
- ▁passiert
- ▁Hand
- fall
- ▁irgendwann
- ▁Problem
- war
- qu
- fühl
- ▁Wer
- ▁wissen
- ▁dort
- ▁jeder
- ca
- ▁deswegen
- sprech
- ▁davon
- ▁damals
- trag
- ▁nämlich
- ▁Punkt
- ▁Welt
- ▁abge
- '7'
- log
- ▁sogar
- ▁kam
- legen
- ▁Moment
- igkeit
- ▁konnte
- ▁komm
- ▁gewesen
- ▁anders
- ▁Bi
- K
- ▁eigene
- ▁liebe
- ▁Teil
- ▁Lo
- ▁toll
- ▁Arbeit
- ▁Seite
- genommen
- ▁to
- ▁alt
- ▁trotzdem
- ▁gehört
- ▁Jetzt
- ▁mache
- ▁Dr
- ▁relativ
- sicht
- ▁steht
- ▁Auto
- ▁darüber
- nehm
- ▁irgendwas
- ▁ohne
- ▁Geld
- ▁Euro
- ieß
- suche
- ▁vier
- einander
- ▁Grund
- ▁Gefühl
- gestellt
- ▁sa
- ativ
- G
- ▁darauf
- I
- ▁All
- ▁Anfang
- ▁darf
- ▁Freund
- ▁direkt
- ▁irgendwo
- ▁letzte
- ▁schlecht
- ▁manchmal
- ▁Bild
- ▁Geschichte
- ▁interessant
- E
- ▁komplett
- ▁Ahnung
- bringen
- nutz
- bild
- ▁frag
- V
- ▁Kind
- ▁meisten
- ▁gehabt
- ▁gedacht
- ▁erstmal
- ▁fast
- ▁stimmt
- '3'
- laufen
- ▁bestimmt
- zahl
- ▁Über
- kommt
- gegangen
- setzen
- ▁funktioniert
- ▁spielen
- ▁Person
- ▁Sinn
- ▁dachte
- ▁fünf
- ▁hoch
- bereit
- ▁brauche
- ▁zwischen
- ▁Spaß
- ▁spannend
- ▁ehrlich
- ▁krass
- ▁schreib
- ▁zumindest
- zeug
- ▁Musik
- W
- fahr
- ▁solche
- ▁Deutschland
- ▁gespielt
- geschrieben
- Ä
- ▁später
- Y
- O
- H
- '!'
- U
- N
- Q
- Ö
- X
- Z
- J
- '%'
- Ü
- é
- «
- »
- '&'
- Ã
- à
- ş
- q
- ¤
- Ÿ
- €
- è
- ı
- ç
- ú
- ë
- ¶
- á
- ć
- —
- õ
- ğ
- í
- °
- ô
- _
- ó
- /
- å
- $
- ́
- û
- ›
- ê
- ‹
- '"'
- ñ
- Ş
- č
- )
- É
- μ
- ø
- š
- о
- ł
- ù
- ã
- ā
- ©
- а
- ':'
- е
- œ
- и
- н
- â
- î
- т
- ń
- р
- к
- 你
- æ
- „
- Č
- с
- ♪
- д
- Š
- в
- ï
- İ
- л
- À
- у
- ь
- я
- м
- ę
- ś
- ž
- п
- '='
- ō
- ř
- Æ
- ш
- з
- ы
- ū
- ș
- Ø
- '~'
- ì
- ò
- ο
- ч
- г
- ý
- ̄
- ц
- Х
- ż
- З
- б
- ¡
- Н
- ă
- ̃
- К
- ж
- ไ
- ồ
- ♫
- ر
- х
- ン
- Ç
- §
- ⁄
- +
- '*'
- Å
- і
- Á
- ī
- џ
- ู
- ;
- '>'
- Î
- ą
- Đ
- Ȗ
- Ε
- έ
- δ
- ι
- λ
- ς
- τ
- υ
- ύ
- О
- Т
- و
- ک
- ں
- ด
- ม
- ่
- ṣ
- “
- ♥
- き
- つ
- ぶ
- ら
- チ
- ッ
- ホ
- ロ
- 中
- 以
- 佢
- 利
- 厲
- 句
- 可
- 吃
- 国
- 士
- 好
- 安
- 害
- 度
- 手
- 晃
- 法
- Ć
- ě
- Б
- ج
- 救
- ά
- –
- ダ
- 制
-
init: null
input_size: null
ctc_conf:
dropout_rate: 0.0
ctc_type: builtin
reduce: true
ignore_nan_grad: null
zero_infinity: true
joint_net_conf: null
use_preprocessor: true
token_type: bpe
bpemodel: data/de_token_list/bpe_unigram1024/bpe.model
non_linguistic_symbols: null
cleaner: null
g2p: null
speech_volume_normalize: null
rir_scp: null
rir_apply_prob: 1.0
noise_scp: null
noise_apply_prob: 1.0
noise_db_range: '13_15'
short_noise_thres: 0.5
frontend: default
frontend_conf:
n_fft: 512
win_length: 400
hop_length: 160
fs: 16k
specaug: specaug
specaug_conf:
apply_time_warp: true
time_warp_window: 5
time_warp_mode: bicubic
apply_freq_mask: true
freq_mask_width_range:
- 0
- 30
num_freq_mask: 2
apply_time_mask: true
time_mask_width_range:
- 0
- 40
num_time_mask: 2
normalize: global_mvn
normalize_conf:
stats_file: exp/asr_stats_raw_de_bpe1024/train/feats_stats.npz
model: espnet
model_conf:
ctc_weight: 0.3
lsm_weight: 0.1
length_normalized_loss: false
preencoder: null
preencoder_conf: {}
encoder: contextual_block_transformer
encoder_conf:
output_size: 256
attention_heads: 8
linear_units: 2048
num_blocks: 22
dropout_rate: 0.1
positional_dropout_rate: 0.1
attention_dropout_rate: 0.0
input_layer: conv2d
normalize_before: true
block_size: 40
hop_size: 16
look_ahead: 16
init_average: true
ctx_pos_enc: true
postencoder: null
postencoder_conf: {}
decoder: transformer
decoder_conf:
attention_heads: 8
linear_units: 2048
num_blocks: 12
dropout_rate: 0.1
positional_dropout_rate: 0.1
self_attention_dropout_rate: 0.0
src_attention_dropout_rate: 0.0
preprocessor: default
preprocessor_conf: {}
required:
- output_dir
- token_list
version: '202211'
distributed: true
```