File size: 1,238 Bytes
d35a7a9
26cafdd
 
 
 
d35a7a9
26cafdd
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
d35a7a9
26cafdd
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
---
datasets: 
  - AliMeeting
language: 
  - zh
license: apache-2.0
metrics: 
  - 
    name: "IHM test CER"
    type: cer
    value: 11.53
  - 
    name: "SDM test CER"
    type: cer
    value: 25.85
  - 
    name: "GSS test CER"
    type: cer
    value: 14.22
tags: 
  - k2
  - icefall
---

# AliMeeting

This is an ASR recipe for the AliMeeting corpus. AMI provides recordings from the speaker's
headset microphones and an 8-channel microphone array.
We pool data in the following 4 ways and train a single model on the pooled data:

(i) individual headset microphone (IHM)
(ii) IHM with simulated reverb
(iii) Single distant microphone (SDM)
(iv) GSS-enhanced array microphones

Speed perturbation and MUSAN noise augmentation are additionally performed on the pooled
data.

## Performance Record

### pruned_transducer_stateless7

The following are decoded using `modified_beam_search`:

| Evaluation set           | eval CER    | test CER |
|--------------------------|------------|---------|
| IHM                      |  9.58  | 11.53 |
| SDM                      |  23.37  | 25.85 |
| MDM (GSS-enhanced)       |  11.82  | 14.22 |

See the [recipe](https://github.com/k2-fsa/icefall/tree/master/egs/alimeeting/ASR_v2) for details.