DSP Innovations - TWELP 1600 bps Vocoder

Details: Written by DSP Innovations expert; Published: 2016 November 09; Last Updated: 2025 February 25

(version 5.0) offers speech quality and intelligibility that were previously achievable only at double or higher bitrates. In our search for a worthy competitor and given the growing demand for free codecs in recent years, we decided to compare our vocoder with CODEC2, which operates at twice the bitrate, to highlight the fundamental difference between a free and a commercial solution.

For Digital HF Radio and other markets.

TWELP Technology Features. The vocoder is based on newest technology of speech coding called "Tri-Wave Excited Linear Prediction" (TWELP) that was developed by experts of DSPINI.

TWELP technology is a new class of vocoders that differs from any other LPC-based vocoders by:

advance reliable method of pitch estimation
pitch-synchronous analysis
advance tri-wave model of excitation
newest quantization schemes
pitch-synchronous synthesis

Thanks to these unique features, TWELP technology provides significantly better speech quality than other well-known technologies—including AMBE+2, MELPe, ACELP, and others—at equivalent bit rates ranging from 300 bps to 4800 bps and beyond.

Additionally, unlike other low-bitrate vocoders (such as MELPe, for example), TWELP delivers much higher quality for non-speech signals, including sirens, background music, and similar audio.

Speech Quality. This is a comparison with the CODEC2 vocoder, which operates at 3200 bps and 1600 bps.
The TWELP 1600 bps and CODEC2 3200 bps and 1600 bps vocoders were tested using the ITU-T P.50 speech base in 20 different languages.

Note:
We have updated the speech database by minimizing inter-speech pauses to eliminate their impact on the evaluation results. Therefore, the numbers obtained from the quality measurements using this updated speech database differ from those previously obtained with the original speech database, where speech pauses were not removed.

The ITU-T P.862 tool was used to evaluate speech quality in terms of PESQ scores:

The diagram demonstrates a substantial difference in speech quality after processing with the TWELP 1600 bps vocoder compared to CODEC2, even when the latter operates at a bitrate twice as high - 3200 bps.. Exact numbers are shown in the table below.

Language	TWELP 1600 bps	CODEC2 3200 bps	CODEC2 1600 bps
American	3.086	2.758	2.550
Arabic	3.040	2.823	2.581
British	2.946	2.497	2.351
Chinese	3.028	2.912	2.658
Danish	3.094	2.794	2.575
Dutch	2.905	2.506	2.333
Finnich	2.886	2.418	2.244
French	3.110	2.849	2.649
German	3.070	2.680	2.529
Greek	3.070	2.768	2.527
Hindi	3.063	2.756	2.538
Hungarian	3.083	2.993	2.747
Italian	3.220	3.029	2.776
Japanese	3.035	2.777	2.607
Norwegian	3.067	2.943	2.710
Polish	3.027	2.668	2.485
Portuguese	3.128	2.911	2.660
Russian	2.977	2.724	2.551
Spanish	3.060	3.096	2.775
Swedish	3.105	2.777	2.590
Average	3.050	2.784	2.572
The average difference between TWELP 1600 bps and CODEC2 3200 bps is 0.266 PESQ, while the average difference between TWELP 1600 bps and CODEC2 1600 bps is 0.478 PESQ.

Speech Intelligibility. Here is a comparison with the CODEC2 vocoder, which operates at twice the bitrate (3200 bps) and at the same bitrate (1600 bps) . The TWELP 1600 bps vocoder and the CODEC2 3200 bps and 1600 bps vocoders were tested using the ITU-T P.50 speech database, covering 20 different languages.
STOI (Short-Time Objective Intelligibility) and ESTOI (Extended Short-Time Objective Intelligibility) metrics were used to assess speech intelligibility:

The diagram demonstrates a significant advantage of the TWELP 1600 bps vocoder over both the CODEC2 1600 bps and 3200 bps vocoders, despite the latter operating at double the bitrate. Exact values are provided in the table below:

Language	TWELP 1600 bps	CODEC2 3200 bps	CODEC2 1600 bps
American	90.61	87.89	84.28
Arabic	89.72	87.31	83.54
British	88.66	84.77	80.45
Chinese	89.66	88.90	85.31
Danish	90.81	87.96	83.65
Dutch	89.54	85.27	81.35
Finnich	86.83	84.70	78.75
French	90.12	88.75	85.10
German	90.20	87.04	82.90
Greek	90.06	88.01	83.02
Hindi	90.09	86.19	82.32
Hungarian	89.62	88.74	85.71
Italian	89.71	87.67	83.25
Japanese	89.94	88.03	84.33
Norwegian	90.82	89.90	86.09
Polish	90.32	87.82	83.63
Portuguese	89.61	89.07	85.28
Russian	89.08	87.01	83.37
Spanish	89.25	88.31	83.22
Swedish	88.55	87.66	83.10
Average	89.66	87.55	83.43
The average difference between the TWELP 1600 bps and the CODEC2 3200 bps vocoders is 2.11%, while the average difference between the TWELP 1600 bps and the CODEC2 1600 bps vocoders is 6.23%.

Considering that a low-bitrate vocoder is a nonlinear device that significantly distorts the spectrum of the original speech signal, the ESTOI metric provides more accurate assessments of speech intelligibility after vocoding:

Language	TWELP 1600 bps	CODEC2 3200 bps	CODEC2 1600 bps
American	82.27	76.98	69.29
Arabic	84.23	76.53	69.49
British	82.30	73.62	64.38
Chinese	84.99	81.45	74.65
Danish	84.81	77.55	69.16
Dutch	83.36	74.75	66.38
Finnich	80.18	75.85	65.07
French	84.40	79.25	71.46
German	83.33	75.19	66.35
Greek	85.05	79.58	70.85
Hindi	83.08	72.58	63.79
Hungarian	83.15	77.81	70.62
Italian	83.73	78.09	70.01
Japanese	85.03	79.01	72.43
Norwegian	85.26	80.47	73.49
Polish	84.68	77.04	68.50
Portuguese	83.82	79.05	71.33
Russian	82.97	76.04	68.74
Spanish	83.70	80.41	71.28
Swedish	80.90	77.17	68.09
Average	83.66	77.42	69.27
The average difference between the TWELP 1600 bps and the CODEC2 3200 bps vocoders is 6.24%, while the average difference between the TWELP 1600 bps and the CODEC2 1600 bps vocoders is 14.39%.

You can download the P.862 and STOI/ESTOI utilities, along with all speech samples, by using the links in the 'Downloads' section at the bottom of the page, and then check all the numbers presented above.

Speech Samples (WAV-files).
A few independent experts compared the TWELP 1600 bps vocoder with the CODEC2 3200 bps and 1600 bps vocoders using the preference method. All listeners preferred the TWELP vocoder at both CODEC2 bitrates, noting that the speech was clearer, sounded more natural, and was less synthetic.

You can play and listen to short samples of the source speech, as well as the speech processed by the CODEC2 3200 bps, 1600 bps vocoders and the TWELP 1600 bps vocoder, using the links in the table below.

For the best listening experience, we recommend using high-quality headphones or premium audio equipment to hear the nuances and differences in the vocoder sound more clearly.

You can also download the complete set of P.50 samples as zip files for all languages simultaneously by using the links in the 'Downloads' section at the bottom of the page.

Language	Source speech	CODEC2 3200 bps	CODEC2 1600 bps	TWELP 1600 bps
American
Arabic
British
Chinese
Danish
Dutch
Finnich
French
German
Greek
Hindi
Hungarian
Italian
Japanese
Norwegian
Polish
Portuguese
Russian
Spanish
Swedish

Superiority In Quality Of The Non-speech Signals. In contrast to other LBR vocoders (MELPe, AMBE+2, etc.), TWELP vocoders provide high quality of non-speech signals, including police, ambulance, fire sirens, etc. This feature in conjunction with high quality natural human-sounding of voice makes TWELP vocoders well suitable for replacement of analog radio by digital radio and also for other applications where high quality transmitting of non-speech signals is relevant along with high quality transmitting of speech signals.

Source signal	CODEC2 3200 bps	CODEC2 1600 bps	TWELP 1600 bps

High Robustness To Acoustic Noise. In contrast to other LBR vocoders, TWELP vocoders are highly robust to acoustic noise due to a reliable pitch estimation method and other features of TWELP technology.

Language	TWELP 1600 bps	CODEC2 3200 bps	CODEC2 1600 bps
American	71.93	66.85	60.47
Arabic	73.60	69.67	63.37
British	69.80	64.23	56.25
Chinese	76.05	72.12	66.55
Danish	71.55	67.72	60.34
Dutch	69.24	63.56	56.34
Finnich	68.27	66.63	57.52
French	73.30	70.15	62.95
German	70.58	66.59	58.97
Greek	73.89	70.33	63.21
Hindi	69.62	62.65	54.79
Hungarian	73.91	69.80	62.71
Italian	70.99	67.33	60.84
Japanese	75.61	70.44	64.33
Norwegian	75.42	71.75	65.99
Polish	72.90	67.06	60.17
Portuguese	72.55	69.92	63.25
Russian	71.20	66.10	59.67
Spanish	74.20	71.54	64.17
Swedish	65.92	64.85	57.36
Average	72.03	67.96	60.96
The average difference between the TWELP 1600 bps and the CODEC2 3200 bps vocoders is 4.07%, while the average difference between the TWELP 1600 bps and the CODEC2 1600 bps vocoders is 11.07%.

Additionally, the TWELP vocoder features an NCSE (Noise Cancellation Speech Enhancement) preprocessor, which cleans the input speech signal from noise and enhances speech quality.
Below, you can listen to a short fragment of heavily noisy English speech after passing through CODEC2 and TWELP vocoders, with NCSE disabled and enabled, respectively.

NCSE	Input speech (SNR=10dB)	CODEC2 3200 bps	CODEC2 1600 bps	TWELP 1600 bps
Disabled
Enabled

The NCSE integrated into the TWELP vocoder is described in more detail on the webpage for our standalone product, 'NCSE-AGC Preprocessor'.

High Robustness To The Channel Errors.

The TWELP technology offers highly efficient speech compression by eliminating redundancy while preserving excellent quality and intelligibility. To enhance robustness against transmission errors, we provide specialized versions called TWELP Robust.
These vocoders are based on an effective Joint Source-Channel Coding approach. Each vocoder is equipped with a custom-designed FEC, tailored to its specific characteristics and operational conditions.
TWELP Robust vocoders provide high speech quality simultaneously in noisy channel as well as in noiseless channel. FEC can operate with "soft decisions" as well as with "hard decisions" from a modem. "Soft decisions" mode provides much better robustness in comparison with the "hard decisions" mode.
For all users of our non-robust vocoder versions, we offer the following recommendations.

The diagram below illustrates the sensitivity of bits at the output of the vocoder to communication channel errors.
Essentially, the diagram shows by what percentage speech quality is reduced when a specific bit is distorted. The first bits in order cause catastrophic distortions, while the latter bits have significantly less impact on quality.

	TWELP 1600 bps Vocoder
1	82.44
2	79.42
3	79.13
4	73.81
5	65.63
6	63.42
7	58.72
8	58.43
9	53.67
10	53.42
11	53.20
12	52.65
13	52.23
14	51.52
15	48.79
16	48.07
17	46.84
18	46.47
19	45.40
20	43.66
21	42.98
22	42.04
23	39.90
24	39.67
25	37.07
26	36.80
27	35.65
28	34.85
29	33.45
30	32.01
31	27.44
32	26.83
33	25.39
34	23.97
35	23.61
36	18.36
37	17.96
38	17.16
39	15.65
40	11.63
41	11.51
42	11.33
43	10.85
44	9.58
45	9.04
46	7.97
47	7.52
48	7.25
49	7.05
50	7.02
51	6.01
52	5.84
53	5.58
54	5.35
55	5.04
56	5.02
57	4.63
58	4.44
59	3.33
60	3.30
61	3.05
62	2.61
63	1.95
64	1.37

We strongly recommend using FEC (Forward Error Correction) with unequal protection of the bits in strong accordance with their sensitivity to errors and utilizing 'Soft Decisions' decoding. This will provide the highest robustness of the vocoder against errors in the channel.

Additional Functionalities. The following additional functionalities are developed by DSPINI and integrated into TWELP vocoders:

Noise Cancellation Speech Enhancement (NCSE)
Automatic Gain Control (AGC),
Voice Activity Detector (VAD),
Discontinuous Transmission (DTX),
Tone Detection/Generation (Single tones and Dual tones). The tones are transmitted by the vocoder facilities.

Note:
The Tone Detector/Generator functionalities are not integrated into the code by default but can be added free of charge upon request.

Each functionality has unique features, performance and characteristics, providing significant superiority over any well-known implementations on the market.

Technical Characteristics And Resource Requirements:

Technical characteristics
Bit Rate (bps)	Algorithm	Frame size (ms)	Algorithmic delay (including frame size) (ms)	Sampling rate (kHz)	Signal format	Bit stream format
1600	TWELP	40	60	8	Linear 16-bit PCM	64

Additional functionalities
Name	Functionality	Technical characteristics
Name	Functionality	Name	Value
AGC	Automatic Gain Control	Control range:	0 ... +42 dB
NCSE	Noise Canceller - Speech Enhancer	SNR increasing	20 dB
NCSE	Noise Canceller - Speech Enhancer	Speech quality improvement	> 0.1 PESQ
Tone Detector	Single/Dual tones detection	In accordance with international standards
Tone Generator	Single/Dual tones generation	Special generator, kept continuity of signal (phase and amplitude of signal of previous frame)
DTX	Discontinuous Transmission	Reduces bit rate down to 110 bps in pauses between active speech regions
VAD	Voice Activity Detection	High reliability even with pink noise at an SNR < 0 dB.
CNG	Comfort Noise Generation	Type of noise	"white"
CNG	Comfort Noise Generation	Level	- 60 dB

The NCSE and AGC integrated into the TWELP vocoder are described in more detail on the webpage for our standalone product, 'NCSE-AGC Preprocessor'.

Resources for ARM Cortex-M4 platform
Module	MIPS* peak	Memory (KBytes)
		Program	Data
		Program	Constants	Channel	Heap	Stack
Voice Encoder	105	36	185	4.6	4.0	1.0
NCSE	6.2
AGC	0.2
Voice Decoder	10.3
Voice Encoder + Voice Decoder	115.3
Total	121.7

Resources for TI's C64 DSP platform
Module	MIPS* peak	Memory (KBytes)
		Program	Data
		Program	Constants	Channel	Heap	Stack
Voice Encoder	38.3	68	185	4.6	4.0	1.0
NCSE	2.8
AGC	0.2
Voice Decoder	4.4
Voice Encoder + Voice Decoder	42.7
Total	45.7

Resources (estimated) for TI's C55 DSP platform
Module	MIPS* peak	Memory (KBytes)
		Program	Data
		Program	Constants	Channel	Heap	Stack
Voice Encoder	64.0	22	185	4.6	4.0	1.0
NCSE	6.5
AGC	0.2
Voice Decoder	10.0
Voice Encoder + Voice Decoder	74.0
Total	80.7

* DSPINI continues optimization of the TWELP algorithm and code in order to minimize computational complexity of the vocoder.

Software Integrity and Security. DSPINI guarantees the ABSOLUTE integrity of its software, free from any undocumented features, undeclared capabilities, or hidden functions. Our customers can be assured that none of our software/code contains any secret features or functionalities concealed from the user. If necessary, we are ready to provide the source code of our software products for appropriate certification.
Moreover, our software is available in source code form—you simply need to purchase the appropriate license to use it.

Guarantee And Support. DSPINI guarantees a quality and accordance of all technical characteristics of the product to requirement of current specifications. Testing and other method of quality control are used for guarantee support.

Any Platforms. DSPINI can port this vocoder software into any other DSP, RISC or general- purposes platform inshort time: 1-2 months.

Licensing Terms. To use the vocoder, customer should obtain a license from DSPINI only.

Customization. The vocoder can be customized under any specific requirements- other bit rate, frame size, any other robustness to channel errors, etc. Please contact with us for details.

Prospects. DSPINI is impoving and developing continuously a set of new vocoders with range from 300 bps up to 9600 bps, based on TWELP technology.

Related Software. This vocoder may be effectively used in a bundle with other DSPINI's products:

Linear and acoustic echo cancellers,
Multichannel noise cancellers (including two-microphone adaptive array),
Wired or radiomodems for any types of channels and bitrates,
Other products.

Downloads:

Datasheet (pdf)
ITU-T P.50 source speech samples (zip)
CODEC2 3200 bps speech samples (zip)
CODEC2 1600 bps speech samples (zip)
TWELP 1600 bps speech samples (zip)
P.862 and STOI/ESTOI utilities
PC-evaluation package (zip) — on request
User's Guide document (pdf) — on request